|J Pathol Inform 2017,
The digital pathology association's annual conference october 1-3,manchester grand hyatt,San Diego, CA
|Date of Web Publication||28-Nov-2017|
Source of Support: None, Conflict of Interest: None
|How to cite this article:|
. The digital pathology association's annual conference october 1-3,manchester grand hyatt,San Diego, CA. J Pathol Inform 2017;8:46
| Oral Abstracts: Interoperability and Component Testing in Whole-Slide Imaging Systems|| |
Esther Abels1, Aldo Badano2
1Emerging Businesses, Philips Digital Pathology Solutions, Hengelo, The Netherlands, 2Division of Imaging, Diagnostics, and Software Reliability, Office of Science and Engineering Laboratories, Center for Devices and Radiological Health, U.S. Food and Drug Administration, Silver Spring, MD, USA. E-mail: [email protected]
A WSI system consists of many components including slide feeders and handlers, image acquisition devices, image processing and transmission devices, display systems and image viewers. After the granting of the de novo for the Philips WSI system, industry and users are now asking about replacing/swapping WSI components, raising issues related to interoperability and regulatory compliance. In this presentation, we will discuss possible regulatory pathways for swapping WSI components, mainly discussing interoperability. We will also review the regulatory history of component and accessory swapping in radiological devices and highlight similarities and differences with respect to digital pathology and WSI devices.
| Workflow Improvements and Challenges in a Pathology Laboratory after Digitalization|| |
1Digital Pathology Team, Laboratorium Pathologie Oost-Nederland, Hengelo, Overijssel, the Netherlands. E-mail: [email protected]
LabPON made the transition to digital diagnosis in July 2015. We believe we are one of the first laboratories to do this, whereby all histological slides are scanned and most of the diagnoses are made digitally. After the transition to digital diagnostics we faced some challenges, but we also found that the quality of our diagnostics and, especially, the logistics in LabPON had improved in many aspects. In 2013, we used a flow analysis to investigate the logistics of digital diagnostics. However, after the transition of our diagnostic facility to fully digital we began setting up a new, more accurate flow analysis. I expect to finish this analysis before May 2017. I have used the results of the current flow analysis, together with all the gained experience, to get a good overview of the challenges and have noted real quality and logistics improvements in digital diagnostics in our laboratory. We found that the digital diagnostic process is faster, more efficient and qualitatively better than the microscope.
| Cloud-Based GPU Computing in Support of Real-time Differential Diagnosis Generation from WSI Data Using Deep Learning|| |
Ulysses G. J. Balis1
1Department of Pathology, University of Michigan Health System, Ann Arbor, Michigan, USA. E-mail: [email protected]
Cloud-based computing and specifically, GPU-enhanced computation, represents a significant incremental opportunity for realization of real-time computation for machine vision tasks that what were previously considered as being non-trivial, including histopathological classification and automated annotation. This presentation will review this rapidly developing field, with an emphasis on approaches that are already computationally feasible. Live demonstrations will be included.
| Towards a Digital Pathology Commons|| |
Michael J. Becich1
1Department of Biomedical Informatics, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA. E-mail: [email protected]
National Institute of Health is focused on creating data commons including funding the NCI's Genome Data Commons to enable and promote data sharing for discovery science and health. The Pathology Imaging community needs to develop infrastructure for a Digital Pathology Data Commons that can contribute to the important clinical and translational opportunities that imaging data sharing provides both clinical care and industry. To take advantage of these major opportunities we need to develop a plan for a Data Commons based on FAIR (F=findable, A=accessible, I=interoperable, R=reusable) which will require deep annotation of images and linkage to other aspects of the health care record. To accomplish this we need to consider the highest value research and teaching data including patient data which could be developed into a large scale sharable repository and “best of breed” information systems which serve the needs of students, researchers and industry partners. The coupling of genomic data and other laboratory data (anatomic, clinical and molecular pathology records) in an innovative fashion should help to establish other best of breed information practices for precision medicine.
| College of American Pathologists Guideline of Quantitative Image Analysis for HER2 Immunohistochemistry for Breast Cancer|| |
1Moffitt Cancer Center, Tampa, Florida, USA. E-mail: [email protected]
Advancements in genomics, computing and imaging technology have spurred new opportunities to use quantitative image analysis for diagnostic testing. Quantitative image analysis (QIA) has been shown to improve accuracy, precision and reproducibility of interpretation than manual scoring by pathologists. Such accuracy is essential to patient diagnosis, prognosis and treatment planning. However, QIA has not gained widespread acceptance. One practical gap is the lack of guidelines in how to perform QIA. The College of American Pathologists (CAP) convened a panel of pathologists and histotechnologists with expertise in digital pathology, immunohistochemistry (IHC), and quality management to develop an evidence-based guideline that provides recommendations in the interpretation of HER2 IHC where QIA is employed. This presentation will provide an update of the guideline including the 11 draft recommendations and the feedback received during the open comment period. The CAP Pathology and Laboratory Quality Center (The Center) develops evidence-based guidelines and consensus statements related to the practice of pathology and laboratory medicine. Through this work, the CAP and its members continually improve the quality of diagnostic medicine and patient outcomes.
| Quantitative Image Analysis of Immununohistochemical Biomarker of Breast Cancer-Practical Issues and a Case-Based Approach following CAP Guidelines|| |
Marilyn M. Bui1, Liron Pantanowitz2, Yan Peng3
1Moffitt Cancer Center, Tampa, Florida, 2Department of Pathology, University of Pittsburgh Medical Center, Pittsburgh, PA, 3Department of Pathology, UT Southwestern Medical Center, Dallas, TX, USA. E-mail: [email protected]
Quantitative image analysis (QIA) has been shown to improve accuracy, precision and reproducibility of interpretation of immunohistochemical (IHC) biomarkers of breast cancer than manual scoring by pathologists, but has not gained widespread acceptance. Recently CAP developed a guideline including 11 recommendations aims to ensure that diagnoses of HER2 IHC for breast cancer are accurate and consistent. This workshop (1.5 hours) is intended to provide the participants a practical and interactive experience to perform QIA of breast cancer biomarkers (estrogen receptor, progesterone receptor, HER2, Ki67, and p53). Before the workshop, the participants will have web access to the whole slide image of cases of invasive breast cancer with biomarker IHC staining and practice on QIA. In the first hour (CME-earning), faculty will discuss commonly encountered issues and offer solutions or recommendations. In the last 30 mins (non-CME-earning), 2 vendors (one with FDA-approved platform and one with non-FDA approved platform) will share with the participants their approach to generate an accurate interpretation. At the end, the answers to the cases will be revealed.
| Hot Topic Discussion: Joint NSH/DPA On-line Digital Pathology Certificate of Completion and Current Certification Environment|| |
Elizabeth A. Chlipala1, Liron Pantanowitz2
1Partner, Premier Laboratory, LLC, Boulder, CO, 2Department of Pathology, University of Pittsburgh Medical Center, Pittsburgh, PA, USA. E-mail: [email protected]
As digital pathology matures it is apparent that we need well trained individuals to manage our digital imaging systems. This panel discussion will introduce the joint NSH/DPA online Digital Pathology Certificate of Completion available in the fall of 2017. An overview of how this program was developed, the content of the educational modules, and the manner in which this program is being provided will be discussed. This session will also include an open dialogue on the certificate of completion program and the current formal certification environment for digital pathology.
| Computational Pathology 101|| |
Francesco Ciompi1, Geert Litjens1
1Radboud University Medical Center, Nijmegen, Gelderland, The Netherlands. E-mail: [email protected]
Computational pathology is a promising field of research that is capable of producing powerful algorithms to aid pathology diagnostics. Expectedly, such algorithms will enter routine diagnostics within the next few years. Especially so-called 'deep learning' techniques have become the standard technique for digital pattern recognition in histopathological images. However close to being of added value, pathologists largely lack knowledge of these techniques and their development process. Targeted audience: pathologists that want to find out what the real benefits are for introducing digital pathology, and researchers in computational pathology.
| DICOM Impact on WSI in Pathology|| |
1DICOM Editor, PixelMed, Bangor, PA, USA. E-mail: [email protected]
As WSI moves from the research realm towards widespread clinical application, and regulatory hurdles are surmounted, the need to assure that digital images are interchangeable and interoperable between sites and vendors becomes paramount. The DICOM Standard was extended in 2010 to support storage of tiled whole slide images, using a pattern of encoding similar to that used by many commercial vendors in their proprietary formats. The use of DICOM allows for the re-use of existing infrastructure (PACS and VNAs) for storage and regurgitation of WSI data, as well as providing standardized protocols for frame (tile) and metadata access for virtual microscopy applications. This presentation will address the advantages and challenges of using the DICOM format, and discuss both technical barriers, such as the absence of support in commonly used open source WSI tool, as well as non-technical obstacles, including intellectual property related issues that have had a dampening effect until recently being resolved. Lessons learned from other specialties that routinely use DICOM, including radiology and cardiology, will be discussed, as well as the similarities and differences between WSI workflow and other specialties. The advantages of leveraging the vast amount of open source and commercial software that already exists for DICOM, as well as the general IT-related advantages of centralizing on a common storage infrastructure throughout the enterprise, particularly for reliability, scalability and security, will be emphasized.
| Implementation of Whole Slide Imaging for Primary Diagnosis|| |
Andrew J. Evans1, Paul J. Van Diest2,3
1Staff Pathologist and Associate Professor, University Health Network, Laboratory Medicine Program, Toronto General Hospital, Toronto, ON, Canada, 2Department of Pathology, University Medical Center Utrecht, Utrecht, The Netherlands, 3Johns Hopkins Oncology Center, Baltimore, MD, USA. E-mail: [email protected]
With the approval of primary diagnosis by the FDA in April 2017, there is renewed excitement and energy centered on the use of whole slide imaging (WSI) for primary diagnosis. As these devices are adopted in the routine laboratory practice, more and more labs are looking for effective solutions for implementing and validating WSI. In addition to selecting a system as well as validation, labs need to be aware of the necessary quality control parameters that will need to be developed and standardized. This session will discuss practical issues that arise when selecting a WSI system, implementing and validating them for primary diagnosis. Three key aspects will be reviewed in this workshop: selection, implementation and validation of the WSI systems. Two experienced adopters from Canada and Netherlands will present their experience with WSI implementation. The work shop moderator and presenters will also share experiences, challenges and lessons with the attendees about implementing WSI for primary diagnosis at their various institutes.
| Introduction of NAGASAKI-NET: Diagnostic and Communication Network to Develop Pathologists among 3 Large Institutes|| |
Junya Fukuoka1,2, Wataru Uegami2, Han-Seung Yoon1, Yukio Kashima3, Kishio Kuroda1, Daiki Isuge2, Oi Harada2, Takashi Hori2, Atsushi Chugo2
1Department of Pathology, Nagasaki University Hospital, Nagasaki, 2Department of Pathology, Kameda Medical Center, Kamogawa, Chiba Prefecture, 3Awaji Medical Center, Sumoto, Japan. E-mail: [email protected]
Objection: Shortage of pathologists especially inside Japan has been considered as a major problem for years. Development of new pathologists, however, takes long time and thus immediate solution is difficult. To solve this issue, we believe, establishing the digital pathology networking and the creation of innovative system of education are important. Recently, we have developed diagnostic and communication network named, NAGASAKI-NET, connecting 3 institutes: Nagasaki University Hospital, Kameda Medical Center, and Awaji Medical Center. We also set a branch office in Tokyo for the expert consultation. Results: The pathologists' cockpits with multiple monitors have completed in all institutes on April 2017. VPN networking to observe WSI from each institute were also completed until then. Pathological LIS data of Kameda Medical Center and Awaji Medical Center became visible through same VPN. EMR and full LIS data of Kameda Medical Center were accessible from Nagasaki University Hospital. WSIs of biopsy and small surgery specimen were shared and signed out using both collaboration tools of WSI viewer and full-time web-based conference system, WebEX. Desktop-sharing was also used when the viewer does not work for collaboration. Sign-out sessions for those cases have been implemented everyday successfully. Conclusion: Educational and diagnostic network, NAGASAKI-NET, has been established among 3 institutes. Here, we propose the importance of education using full time web-communication tool and WSI collaboration through VPN.
| Are we there yet? Quantitative Multiplexed Tissue Assay Development for Clinical Trial Support|| |
1Research Pathologist, Genentech, Inc., San Francisco, CA, USA. E-mail: [email protected]
Immunohistochemistry (IHC) is still a mainstay in the industry-sponsored clinical trials, however results have been limited to qualitative or semi-quantitative results at best - usually answering only the question “is the target present”? In order for high-order multiplexing technologies to advance out of academic laboratories and into prospective clinical trials, we must offer more utility than other quantitative assays. Supporting immuno-oncology programs has created opportunities for digital pathology and multiplex immunoassays to move into clinical trials in two ways: (1) we can use less tissue by combining analytes on one slide, and (2) we can enumerate and localize cellular populations of the tumor microenvironment in a way that other quantitative assays cannot. In the first case, programs of targeted combination therapies aim to drive the immune system into the tumor while simultaneously turning off the regulatory brakes. In order to confirm the mechanism of action, on-treatment needle core biopsies are necessary but provide very limited tissue for pharmacodynamic assays. In addition, multiplexed IHC can subset immune cells and measure their spatial relationship with each other and tumor cells, potentially providing answers to improve patient selection. As these assays move into the prospective clinical space, it will be important to overcome current technical challenges and establish a standardized approach to assay development and digital pathology algorithm validation.
| In Vivo Microscopy and its Role in the Future of Pathology|| |
1Pathologist and Researcher, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA. E-mail: [email protected]
In vivo microscopy (IVM) provides microscopic optical images in patients, in real time, without tissue removal. This has tremendous potential to impact patient care, including in vivo disease detection, biopsy guidance, diagnosis, and resection margin assessment, in a variety of organ systems and disease types. Due to its microscopic resolution, IVM has close parallels with traditional histology, and requires knowledge of the histomorphologic features of many disease entities, which is a skill set inherent to pathologists. Therefore, IVM is likely to have major impacts on the way we practice as pathologists. In this talk, we will: 1) Explore current efforts in IVM in a variety of clinical applications, including the GI tract, lung, skin, and breast. 2) Discuss the impacts IVM could have on the practice of pathology in the short-term and long-term future. 3) Discuss the role of pathologists in validating, implementing and interpreting IVM as a complement to traditional tissue pathology.
| A Web-based, Open-source Data Management Platform Supporting Visualization and Analysis Plugins for Deep Learning in Diagnostic Pathology|| |
Charles Law1, Beverly Faulkner-Jones2
1Co-founder, Kitware, Inc., Clifton Park, NY, 2Department of Pathology, Beth Israel Deaconess Medical Center, Boston, MA, USA. E-mail: [email protected]
Background: Digital pathology enables computer analysis for pathology diagnosis. Deep learning is one promising approach, but a major challenge is annotating and managing the thousands of high-resolution images needed for training. Methods: We train with an archive of teaching images (>10,000), and developed an open-source web-application called Girder to annotate and manage the training data. Girder uses python to directly interface with deep learning packages. Our Theano neural network has 8 convolutional and 2 pooling layers with effective input receptive fields of 92x92 pixels. A bootstrap approach to annotation simplifies generation of ground-truth data. New slides are initially processed with the neural network, and the results are proofread by experts using a girder plugin to verify detection results. The network is retrained with this additional data to improve its performance. Generating ground truth becomes easier as the accuracy of the detector improves. Results: Our system has been demonstrated by detecting melanoma cells in lymph nodes, with over 30,000 melanoma cells annotated. Detection accuracy is high, but our results are suspect because test images are partially selected by the network being evaluated. We are in the process of manually annotating an independent set of cells to get results that can be reported without reservations. Conclusions: Deep learning provides a powerful new tool for digital pathology analysis. The deep-learning resources we have developed as plugins to Girder will simplify the process of training networks and open this technology to a wider user base.
| Applications of Digital Pathology in Pathology Education at the Undergraduate and Graduate Level|| |
1Professor of Microbiology and Pathology, New York College of Podiatric Medicine, New York, NY, USA. E-mail: [email protected]
Background: As more and more medical schools move away from the use of the microscope as a means of teaching and studying pathology in favor of virtual microscopy (digital pathology), much has been done to capitalize on the advantages and opportunities of the digital world. This presentation will summarize some of the work done at NYCPM and elsewhere to advance pathology education. Methods: 1) Accessing images from multiple locations to enhance the local slide collection; 2) using available software to annotate slides; 3) archiving strategies for preserving and sharing such “value-added” images. 4) applying data mining strategies to access slides for Tumor Board use and resident training; 5) using a Virtual Diagnostic Lab, accessing digital images in a scenario-driven format, to train residents in slide analysis and staining techniques. Results: A number of websites have been created to make these ideas available to educators world-wide. Feedback will be discussed, where available. Conclusions: The addition of digital capabilities to the pathology classroom has greatly enhanced the study of this discipline. It will aid in the training of professional diagnostic pathologists and of those will use their services in patient care.
| Bringing Digital Pathology to the CLIA laboratory: the Moffitt Cancer Center Experience|| |
Anthony M. Magliocco1
1Department of Anatomic Pathology, Moffitt Cancer Center, Tampa, FL, USA. E-mail: [email protected]
Digital pathology is an exciting suite of technologies and applications with potential to transform the practice of anatomic pathology. We will review the Moffitt Cancer Center experience with Digital pathology including CLIA applications with a focus on opportunities and challenges. There will be discussions of utilization in gross room, education, Quality control, telepathology and advanced image analysis with specific examples and applications for biomarker analysis. There will be discussion of different technologies including single plex, multiplex and high dimensional mass cytometry imaging.
| Poised for Disruption? The Door Opens for Digital Pathology and Whole Slide Imaging|| |
1Editor-In-Chief, The Dark Report, Spicewood, TX, USA. E-mail: [email protected]
It's been almost 20 years since the first telemicroscopy and digital pathology solutions began to appear in the marketplace. There has been plenty of progress during the past two decades, but the regulatory barrier is widely-believed to have retarded wider adoption of digital pathology systems by most anatomic pathology laboratories. The big development was the FDA's clearance of the first digital pathology system for use in primary diagnosis in the United States last April. Proponents of digital pathology are enthusiastic that this important regulatory step will encourage faster and wider adoption of digital pathology systems and whole slide imaging.
| From Diagnostics to Quality Control: The Amazing Possibilities of Artificial Intelligence in the Digital Pathology Setting|| |
1Director, United Kingdom National External Quality Assessment Service, London, 2Pathology, University College London, Academic Pathology, London, Middlesex, UK. E-mail: [email protected]
The UK National External Quality Assessment Scheme for Immunohistochemistry and In-situ Hybridisation was established some 30 years ago with the sole aim of ensuring that as IHC, as it was gradually being embraced in the slide based clinical pathology setting as an aid to diagnosis, it was performed correctly. Essentially UKNEQAS had to assist with the education of those who were carrying out the IHC staining in the early years as there was no real automation or indeed training available to help with developing the IHC service.
Several decades later much in the area of IHC has changed. We now have robust automation, some wonderful reagents and numerous experienced practitioners around the world. There are many thousands of publications on the subject and almost every histopathologist is very experienced in using IHC within their clinical practice. Nonetheless, amongst all this amazing development, significant challenges have arisen. Companion diagnostics such as IHC for Estrogen Receptors, Her-2, Alk and PD-L1 are now with us. These tests can often be very challenging to perform and to interpret. It goes without saying that if these assays are not performed and interpreted correctly then it is the patient who suffers and a premature death may follow.
Her-2 IHC testing in breast cancer was performed very poorly across the world for a number of years, thus denying patients appropriate treatment for their tumours. It was an unsatisfactory state of affairs and it fell to the EQA schemes to try to sort out the mess. UKNEQAS-IHC&ISH did just that, but the achievement required many years of effort and huge numbers the circulation and assessment of great numbers of FFPE sections. The capacity of the experts who performed the central review was stretched to the limit and as a consequence the occasional error was made. Although these were corrected through an “appeals process” it has become clear that it would be so much better if novel analytical technology could be developed to provide a more consistent and reproducible assessment of the UKNEQAS slides that were stained by participants.
After collaboration with others, prototype software for analysis of digitised HER-2 slides has been developed. From the EQA perspective it is showing great promise with regard to assessing the quality of HER2 IHC staining. This same software is being developed to assist with the interpretation of PD-L1 from the EQA aspect too.
| Digital Pathology and Image Analysis in Translational Medicine and Immuno-oncology|| |
Michael C. Montalto1
1Executive Director and Head of Translational Pathology and Clinical Biomarker Laboratories in Translational Medicine, Bristol Myers Squibb, Princeton, NJ, USA. E-mail: [email protected]
Digital pathology and image analysis plays an important role in drug development including tumor profiling, mechanisms of action, and exploratory patient selection. In immuno-oncology, understanding cell phenotypes in the context of the tumor micro-environment is critical and image analysis-based multiplexing can further shed light on the immune-biology of cancer. Beyond the exploratory role of image analysis, clinical platforms for digital pathology are being established. This creates a practical path to transition image analysis-based biomarkers to companion diagnostics. This talk will examine the state of the art of pathology based image analysis in translational immune-oncology and explore the potential of this technology as a companion diagnostic platform.
| Stimulated Raman Scattering Microscopy and Histology of Neurosurgical Specimens|| |
Daniel A. Orringer1, Balaji Pandian1, Yashar S. Niknafs1, Todd C. Hollon1, Julianne Boyle1, Spencer Lewis1, Mia Garrard1, Shawn L. Hervey-Jumper1, Hugh J. L. Garton1, Cormac O. Maher1, Jason A. Heth1, Oren Sagher1, D. Andrew Wilkinson1, Matija Snuderl2,3, Sriram Venneti4, Shakti H. Ramkissoon5,6, Kathryn A. McFadden4, Amanda Fisher-Hubbard4, Andrew P. Lieberman4, Timothy D. Johnson7, X. Sunney Xie8, Jay K. Trautman9, Christian W. Freudiger9, Sandra Camelo-Piragua4
Stimulated Raman scattering (SRS) is a fiber-laser-based microscopy technique that images fresh, unprocessed and unlabeled tissue. We use Raman spectra that detects lipids and proteins/DNA to develop cytological contrast. We have imaged more than 400 intraoperative neurosurgical specimens at the University of Michigan using SRS. We have developed an image processing technique that produces virtual images with a color scheme similar to standard hematoxylin and eosin (HE) staining, in a process we term stimulated Raman histology (SRH). In order to determine the reliability of tissue diagnosis using SRH, an online survey was given to different neuropathologist using both SRH and standard intraoperative HE scanned images in a subset of 30 cases. Diagnostic accuracy was greater than 92% regardless of the imaging technique. Although the coloring in SRH does not correspond to the acid and basic moieties of HE, there is a strong overlap in the cytological and architectural features of the tissue that unable pathologist to render an intraoperative diagnosis of neurosurgical specimens. SRH provides images comparable to standard HE, it is a faster process than standard intraoperative consultation, images can be read and access via any web-based computer, tissue used for SRH is suitable for further histopathological or molecular testing. In the future SRH could be a helpful intraoperative imaging tool, particularly in centers that lack the onsite expertise of a neuropathologist.
| Light-sheet Microscopy for Slide-Free 3D Digital Pathology|| |
1Department of Pathology, University of Washington, Seattle, WA, USA. E-mail: [email protected]
Background: Light-sheet microscopy (LSM) is a fluorescence microscopy technique that produces high-resolution 3D images. However, commercially available LSM systems are ill-suited for clinical applications. Our group fabricated an open-top LSM specifically designed for imaging clinical specimens, i.e. biopsies and surgical resections. We will give an overview of the technology and describe the results for imaging the surface of large fresh specimens and 3D imaging of biopsies. Methods: Tissue was either imaged fresh or chemically clarified by immersion in 2, 2' thiodethanol for 15 minutes, stained with nuclear and cytoplasmic fluorescent stains for 2 minutes, and imaged using our custom-built LSM. The images were digitally reconstructed and false-colored to simulate H and E-staining. Results: Fresh tissues were imaged at high speed (15-30 sec/cm^2) over a wide-area with high resolution (~2 microns). After clarification, core-needle biopsies could be imaged throughout the entire 1 mm thickness. Clinical validation studies show high accuracy (>90%) and added diagnostic value from the 3D information. H and E slides were not affected by fluorescent dye staining, clarification, or LSM imaging. Conclusion: LSM is a slide-free, non-destructive method for acquiring digital pathology images, with the added benefit of providing superior depth of imaging for 3D visualizations. A substantial challenge is the processing and visualization of terabytes of raw imaging data, an area of active investigation in our group. LSM is an attractive technology for slide-free digital pathology, providing strong economic incentives by minimizing laboratory workload while generating rich data for machine learning and 3D visualization.
| Twenty Years of Whole Slide Imaging: The Coming Phase Change|| |
1Department of Biomedical Informatics, SUNY Stony Brook University, Stony Brook, NY, USA. E-mail: [email protected]
I will survey the development of Digital Pathology methodology beginning with the 1997 virtual microscope prototype at Hopkins (PMC2233368) to current tools, methods and algorithms designed to display, analyze and classify whole slide imaging data. I will describe the capabilities of current methods, describe how these methods are likely to evolve and how they will be likely to impact Pathology research and practice.
| Image Analysis Vendor Challenge Workshop|| |
Liron Pantanowitz1, Jeroen van der Laak2
1Department of Pathology, University of Pittsburgh Medical Center, Pittsburgh, PA, USA, 2Department of Pathology, Radboud University Medical Center, Nijmegen, The Netherlands. E-mail: [email protected]
Objective and accurate biomarker assessment is a critical part of many clinical trials. Also, there is an increasing use of such biomarkers for personalized medicine in clinical settings. An example is PD-L1 immunohistochemical staining of tissue sections. The combination of whole slide imaging and image analysis software has been shown to be a powerful means to quantify expression of these markers. Currently there are a number of commercial software solutions available for this, each with its own benefits. This workshop will challenge three vendors who offer state-of-the-art solutions (Indica Labs, Visiopharm and Leica) to showcase their software in a collegial and non-competitive way. Their assignment in the workshop is to demonstrate to the audience how their product would help a client (e.g. a researcher) make an algorithm with their software (or toolkit) to analyze PD-L1 immunohistochemical staining of tissue sections. The aim of this workshop is to highlight possibilities and address potential limitations of image analysis, especially with regard to PD-L1. After the vendor presentations there will be ample time for Q and A.
| Whole Slide Imaging: Quantitative Interpretation in the Clinical Setting|| |
Zoya Volynskaya1, Ozgur Mete1, Sylvia L. Asa1
1Department of Pathology Laboratory Medicine Program, University Health Network, University of Toronto, Toronto, Ontario, Canada. E-mail: [email protected]
Background: Proliferation markers, especially Ki-67, are increasingly important in diagnosis and prognosis. The best method for calculating Ki-67 is still the subject of debate. Methods: We evaluated image analysis tools to provide quantitative robust interpretation of Ki-67 in neuroendocrine tumors. Results: We expanded our primary digital pathology diagnosis to include quantitative interpretation of Ki67 utilizing the Aperio/Leica nuclear algorithm. Slides are digitized during routine scanning for primary diagnosis and accessible through the Cerner CoPath LIS interfaced with Aperio eSlideManager. The algorithm provides robust results that are obtained within seconds, compared to 10-40 min of manual counting that is subjective. As not all cases are scanned routinely, this application increased scanning requests and the number of scanned cases. The initial response from pathologists is extremely positive and has led other subspecialty groups to consider adoption of similar workflow. Adoption was dependent on three major factors: (1) availability of reliable and user-friendly technology and workflow integration for pathologists and histologists, (2) a clear need for the solution and (3) a perceived benefit to the users. The improved technology and ease of workflow increased the number of users and use cases; technologic failures resulted in reduced participation. The benefits identified by users included time savings, ease of viewing, accuracy of annotation and academic benefit. Conclusions: Full adoption of digital pathology provided the infrastructure for implementation of quantitative analysis that is faster and provides accurate, reproducible data. Future improvement will involve automated incorporation of image analysis results into diagnostic reports generated by the LIS.
| Digital Breast Pathology in the NHS: Experience from an Innovative Validation and Training Pilot in the United Kingdom|| |
Bethany Williams1, Darren Treanor1
1Digital Pathology, Leeds Teaching Hospitals NHS Trust, University of Leeds, Leeds, Yorkshire, UK. E-mail: [email protected]
Background: Diagnostic pathology laboratories are increasingly interested in using digital pathology to improve service quality. Safe and successful use of digital microscopy in routine primary diagnosis is reliant on individual pathologists receiving adequate training and gaining sufficient experience of digital cases. We designed an innovative training and validation protocol to allow our pathologists to gain competence and confidence in live digital diagnosis, in a real world, risk mitigated environment. Methods: Four breast pathologists were recruited. They received basic digital microscope training, and completed a set of training cases, compiled to reflect the breadth and depth of breast histopathology, and provide exposure to challenging digital diagnostic scenarios. Following discussion of the training set, pathologists commenced live digital reporting. All breast diagnoses were made on digital slides, with immediate reconciliation with the glass slides before sign-out. Data on the diagnosis, digital:glass discrepancies, diagnostic confidence, and diagnostic modality preference were made for each case. Regular meetings were held to review cases where discrepancies were recorded. After accumulating approximately 2 months experience of primary digital diagnosis with glass review, a validation document was produced for each pathologist, summarizing the cases viewed, overall concordance rates, overall preference for digital or glass diagnosis, and any potential pitfalls for that individual pathologist on the digital microscope. Results: In the initial training phase, digital:glass concordance ranged between 80 and 100% for the four pathologists. Areas of diagnosis implicated in digital:glass discordance were the mitotic count component of invasive tumour grading, and the identification of lymph node micrometastases. In the clinical reporting phase, over 600 live cases were reported digitally, with glass verification prior to sign out. Absolute concordance between digital and glass diagnoses was greater than 95%. On analysis, none of the discordant diagnoses on the digital microscope would have resulted in significant patient harm, and all concerned areas of difficult diagnosis prone to intraobserver variation using any diagnostic medium. Conclusions: Our pilot validation study demonstrates the importance of individualised training and validation for histopathologists in specialty specific primary digital diagnosis. Our pathologists all reported high rates of satisfaction with digital microscopy, and all now report breast cases digitally as standard. Informed by the data from this pilot, full deployment of digital pathology for primary diagnosis in all histopathology subspecialties is under way at Leeds Teaching Hospitals NHS Trust.
| Digital Pathology Using Deep Learning|| |
1Department of Biomedical Engineering, University of Florida, Gainesville, FL, USA. Email: [email protected]
Background: Digital pathology image analysis can provide crucial quantitative support for improving characterizations of various diseases such as lung cancer, breast cancer, brain tumor, etc. However, due to the ever-increasing amount of image data, it is inefficient or even impossible to manually process the image data. Recently deep learning has attracted a large amount of attention in biomedical image analysis area. In this presentation, we will provide a snapshot of multiple applications in digital pathology based on deep learning, specifically for the general cell detection task on various types of pathology images. Methods: Deep learning methods typically take in raw input data and gradually learn hierarchical feature representations via a stack of non-linear transformations. Both theory and practice have proven that deep architecture is capable of learning highly complex functions and uncovering the intrinsic structure in the data. Convolutional neural network (CNN) is one of the most successful and widely used architectures. In this presentation, we propose several CNN based learning frameworks including fully residual convolutional neural network (FCN), deep voting, and structured regression for efficient and robust cell detection. All the three architectures utilize convolutional layers and sub-sampling layers to extract hierarchical feature representations, de-convolutional layers are used in FCN for up sampling the feature maps, which is crucial for end-to-end training. Specifically, FCN does not invoke sliding window testing and is highly suitable for large scale image analysis. Result: We have extensively evaluated our method using phase contrast microscopy images of HeLa cervical cancer, Ki-67 stained Neuroendocrine Tumor (NET), and H and E stained breast cancer, lung cancer, brain tumor, bone marrow, and skeleton muscle microscopy images. In total seven completely different datasets covering different staining methods and image acquisition techniques are used to demonstrate the effectiveness and generalization capability of our proposed deep learning based high throughput image analysis methods. Conclusion: We propose several efficient and general frameworks to rapidly process large scale digitized pathology specimens. They are designed to handle the pathology images that are challenging due to different image modalities, touching cells, background noises, and large variations in cell morphologies. Experimental results demonstrate the robustness and efficiency of the deep learning based methods in analyzing digital pathology specimens.
| Poster Abstracts: Use of Digital Pathology to Drive Revenue to Labs|| |
Dan Angress1,2, Theresa A. Feeser1,3, Michael N. Kent1,4
1DLCS-Clearpath, LLC, Dayton, 4Dermatology Department, Boonshoft School of Medicine, Wright State University, Dayton, OH, 2President, Angress Consulting, Manhattan Beach, CA, 3Paragon Consulting Services, LLC, Saint Petersburg, FL, USA. E-mail: [email protected]
This presentation is geared to the practical use of DP as tool to drive lab revenue, It is intended to show how a regional dermpath lab, DLCS used a DP product, Clearpath to bring on over $1M in additional revenue in less than a year by implementing a remote tool to allow derms to read their slides. THe presentation is not intended to promote Clearpath but simply as a representative application not seen in the industry yet.
| Digital Pathology Workflows for Large-Scale Biomarker Studies|| |
Timothy Baradet1, Vipul Baxi1, George Lee1, Cyrus Hedvat2
1Translational Bioinformatics, Bristol-Myers Squibb, Lawrenceville, 2Translational Medicine, Bristol-Myers Squibb, Lawrenceville, NJ, USA E-mail: [email protected]
Background: This poster will present strategies and tactics for applying image analysis to a large-scale biomarkers studies in involving hundreds or even thousands of whole-slide images. Topics covered: Planning your study Image prep and Scanning Image Analysis and Database Integration Optimizing Workflow QC and Pathology Support Data Analysis. Methods: Serial sections of 25 cases in 8 tumor types were analyzed for 6 biomarkers; a total of 1200 whole-slide images. Following IHC staining and scanning, images were organized in a database and scored using HALO software (Indica Labs). Tumor/stroma classification was achieved on slides with appropriate tumor markers and transferred to serial sections. Successive rounds of pathology QC ensured accuracy for classification and IHC scoring. Data was analyzed and graphed using MATLAB software. Results: High quality biomarker data from large scale whole-slide studies can be collected and analyzed efficiently with organized system of data collection, pathology support and image analysis. Conclusions: Careful planning of workflow, pathology support and data analysis greatly enhance success of large-scale biomarker studies using whole-slide images
| An Automated Approach to Calculate Ki-67 Index in Pancreatic Neuroendocrine Tumors|| |
Bernadette M. Boac1, Daryoush Saeed-Vafa1, Anthony M. Magliocco1, Barbara A. Centeno1
1Department of Pathology, Moffitt Cancer Center and Research Institute, Tampa, FL, USA. E-mail: [email protected]
Background: Pancreatic neuroendocrine tumors (PanNET) account for 3-5% of all pancreatic malignancies and have 5-year survival rates of 42%. The World Health Organization recently adopted a three-tier grading system for PanNETs based on the Ki-67 index. However, these counts are time consuming with poor interobserver reproducibility. Here we test the feasibility of an automated approach to calculate the Ki-67 index via digital image analysis (DIA). Methods: Twenty-Eight PanNET cases (21 G1, 5 G2, 2 G3) were identified at Moffitt Cancer Center. Ki-67 expression was studied via a dual-immunofluorescence assay using synaptophysin. An automated Ki-67 index was calculated via HALO® on the whole slide image (WSI) and 20x-hotspots (averaging three hotspots) and then compared to those in the pathology report. Results: The Ki-67 index calculated via WSI DIA correlated 26/28 times with the manual count, and 27/28 times via 20x-hotspot analysis. Average time for WSI DIA was 9.16 minutes, and was negligible for 20x-hotspot analysis. There was no statistical significance comparing the Ki-67 index calculated via WSI versus 20x-hotspot analysis (p= 0.67), means of 3.49% and 4.86% respectively. The Ki-67 index calculated by DIA was strongly correlated with the pathology report (r2= 0.86, p<0.00001, irrespective of method). Conclusions: This pilot study suggests that DIA could be useful to aid in the grading of PanNETs in that it might improve the reproducibility, reliability, and time of the diagnostic process. It is possible that WSI counting may better predict prognosis in PanNETs. Our results are promising and should encourage further, larger, validation studies.
| Three-Dimensional Imaging, Scanners, and Additive Manufacturing: Applications for Pathology|| |
Navid Farahani1, Hameed Tasal1, Dylan Jutt1, Alex Braun1, Todd Huffman1, Liron Pantanowitz2
1Computational Pathology, 3Scan, Inc., San Francisco, CA, 2Department of Pathology, University of Pittsburgh Medical Center, Pittsburgh, PA, USA. E-mail: [email protected]
Imaging is vital for the assessment of physiologic and phenotypic details. In the past, biomedical imaging was heavily reliant on analog, low-throughput methods, which would produce two-dimensional (2D) images; however, newer digital and high-throughput 3D imaging methods, which rely on computer vision and computer graphics, are transforming the way biomedical professionals practice. 3D imaging has demonstrated value as a resource for both biomedical and medical professionals, aiding in diagnostic, prognostic, and therapeutic decision-making. Herein we summarize current imaging methods that enable optimal 3D histopathologic reconstruction: whole-slide imaging (WSI) scanning and 3D scanning. Emerging automated 3D microscopy platforms, which combine robotics, sectioning, and imaging, aim to digitize entire microscopy workflows and are briefly mentioned. Lastly, both current and emerging 3D imaging methods are discussed in relation to current and future applications within the context of pathology.
| Evaluating Image Analysis Approaches towards “Harmonization” of PD-L1 Assays|| |
Allison S. Harney1, Staci J. Kearney1, Carsten Schnatwinkel1, Famke Aeffner1, Luke Pratte1, Zach Pollack1, Jenifer Caldara1, Karen Ryall1, Joseph Krueger1, Daniel Rudmann1, Roberto Gianani1
1Flagship Biosciences, Inc., Westminster, CO, USA. E-mail: [email protected]
Background: Five different companion or complementary diagnostic tests exist for therapeutics targeting the PD1/PD-L1 pathway. The FDA-AACR-ASCO “PD-L1 Blueprint” working group identified similarities, and differences, between four of the commercialized assays. Each test uses a different interpretation, creating a highly complex diagnostic landscape. We used the Computational Tissue Analysis cTA™ platform to develop a strategy towards normalizing data produced from four PD-L1 diagnostic tests. Methods: Non-small cell lung carcinoma (NSCLC) patient samples were stained with the FDA approved Dako 28-8 and Dako 22C3 tests, and in-house SP142 and E1L3N assays. The cTA™ platform was used to identify tissue and cell-specific Biofeatures™ and then generated scores for PD-L1 test comparison. Results: Comparison of manual and digital scores using cTA™ demonstrated that cTA™ significantly reduces inter-pathologist variability in PD-L1 scoring, as observed by the median percent coefficient of variance decreased from 125% for manual pathology scoring of 3 pathologists to 8% with cTA™. Digital quantification of membrane staining intensity in the tumor compartment using cTA™ show that the average intensity of the 22C3 and 28-8 assays were similar, while SP142 was lower, and E1L3N the highest. The percentage of PD-L1 positive cells identified in each assay was highly correlated across the reference range of PD-L1 expression for each assay. Conclusions: When a single cTA™-based scoring system is applied, the PD-L1 assays are mathematically normalized, harmonizing the scoring. Based on the proof-of-concept demonstrated in this study, a cTA™ approach could enable harmonization of the PD-L1 tests through use of a digital pathology platform.
| Application of Microscope-Based Scanning Software (Panoptiq) for the Interpretation of Cervicovaginal Cytology Specimens|| |
Ruben Groen1,2, Kuniko Abe1, Han-Seung Yoon1, Zaibo Li4, Rulong Shen4, Akira Yoshikawa1, Takao Nitanda1, Yukiko Shimizu3, Isao Otsuka2, Junya Fukuoka1,2
1Nagasaki Educational and Diagnostic Center of Pathology, Nagasaki University Hospital, Nagasaki, Departments of 2Pathology and 3Obstetrics and Gynaecology, Kameda Medical Center, Kamogawa, Chiba, Japan, 4The Ohio State University Wexner Medical Center, Columbus, OH, USA. E-mail: [email protected]
Background: Digital pathology has been increasingly gaining the attention of pathologists worldwide. However, the application of digital cytology is relatively unexplored. The microscope-based scanning software, Panoptiq™, enables the operator to combine low-power panoramic digital images with high-power Z-stacks at regions of interest with a significantly smaller image size than that obtained by whole slide scanning. This study aimed to evaluate the feasibility of the use of Panoptiq™ in the digital interpretation of cervicovaginal cytology specimens in comparison with the conventional light microscope. Methods: One hundred liquid-based cytology slides were selected sequentially. The dotted slides were reviewed and scanned where all dotted areas were further scanned by a high-power objective with Z-stacks. The cases were reviewed by two pathologists and a cytotechnologist using conventional light microscopy and digital cytology images acquired by Panoptiq™, based on the Bethesda classification system. The washout time was set as more than 2 weeks. The Cohen's kappa coefficient was calculated to measure the agreement between the two modalities. Results: Digital cytology showed an inter-modality agreement of two observers who had sufficient training in digital pathology at concordance rates between 88 and 89% with Kappa values between 0.84 and 0.85 while the other observer who did not have sufficient training in digital pathology had lower agreement at a concordance rate of 56% with a Kappa value of 0.40. Conclusion: The microscope-based scanning software, Panoptiq™, is feasible for interpretation of cytology specimens but requires adequate training in digital pathology.
| Preparing for High Throughput Image Analysis|| |
Douglas J. Hartman1
1Department of Pathology, University of Pittsburgh Medical Center, Pittsburgh, PA, USA. E-mail: [email protected]
There are many more opportunities to use image analysis in clinical practice. However, the effective integration of image analysis will require transformation within the pathology laboratory. There are many practical issues that must be considered when building image analysis for routine clinical practice - such as barcoding of immunohistochemical slides, separation of control tissue from specimen tissue, and individual tissue concerns. In this talk, I will discuss the challenges and opportunities as well as preparations necessary in order to implement high throughput image analysis.
| Buying a Digital Pathology System: What Labs and Vendors Need to Know|| |
Douglas J. Hartman1
1Department of Pathology, University of Pittsburgh Medical Center; Pittsburgh, PA, USA. E-mail: [email protected]
With the anticipated approval of digital imaging for primary diagnosis by the Food and Drug Administration, more pathology departments will be requesting demonstrations of digital pathology systems. In order to determine the best solution for a lab, it is important for pathology labs to have a clear understanding of their lab workflow and it is important that vendors are aware of the common pathology workflows. In this talk, I will discuss the workflow within our lab and other common pathology departmental structures within the United States. This talk will be directed to both the pathology lab and vendor community.
| Automated assessment of Ki-67 Expression in Breast Cancers: The Utility of Virtual Triple Staining|| |
Akira I. Hida1,4, Lars Pedersen2, Dzenita Omanovic6, Takashi Ogura3, Yumi Oshiro1, Akihide Tanimoto4, Naoki Kanomata5, Takuya Moriya5
1Department of Pathology, Matsuyama Red Cross Hospital, Ehime, 3Life Science Business, Business Promotion 1st Department, System Division, Hamamatsu Photonics K.K. Hamamatsu, 4Department of Pathology, Field of Oncology, Kagoshima University, Kagoshima, 5Department of Pathology 2, Kawasaki Medical School, Kurashiki, Japan, 2Sales, 6Research and Regulatory Development, Visiopharm, Hørsholm, Denmark. E-mail: [email protected]
Background: Breast cancer (BC) is the most common cancer in women. Highly proliferative tumors with high Ki-67 expression are associated with a poor prognosis, and chemotherapy is recommended. However, there remains confusion regarding the best way to evaluate Ki-67. Counting 1000 cells represents a huge task for pathologists and raises concerns over reproducibility. Methods: We performed digital image analysis (DIA) using Virtual Triple Staining (VTS), which comprises part of the Visiopharm Oncotopix software. Resected 152 BC samples were stained for Ki-67, pan-cytokeratin, and p63 (Autostainer-Link48, Dako). Scanned images (NanoZoomer-XT, Hamamatsu Photonics) were aligned for invasive cancer detection and Ki-67 counting. The algorithm was trained in 29 cases, and validated in the remaining 123 cases. One pathologist manually counted many cancer cells at the Ki-67 slides for labeling index (LI). Results: DIA values were outliers for two cases in the training set, including one with p63-positive staining on cancer cells, and one that was too dense to be counted correctly. The remaining 27 cases showed a strong correlation between the DIA value and LI (Spearman's rank correlation: R=0.88). The median LI and DIA values in the validation set were 23.5 and 24.1, respectively. The DIA process discarded 9 cases as the software under-detected negative nuclei. The rest 114 showed a positive correlation (R=0.75, p<0.0001). Conclusions: Automated DIA using VTS generates reliable results, and is also expected to show high reproducibility. Although some cases still require manual handling by pathologists, DIA is a candidate standard method for assessing Ki-67 expression in pathology.
| First person shooter video game principles enable rapid generation of ground truth for deep learning|| |
Richard Huang1, Jack Zeineh1,2, Marcel Prastawa1,2, Giovanni Koll1,2, Gerardo Fernandez1,2
1Department of Pathology, Icahn School of Medicine at Mount Sinai, 2Center for Computational and Systems Pathology, Icahn School of Medicine at Mount Sinai, New York, NY, USA. E-mail: [email protected]
Background: Deep learning has the potential to change pathology through the creation of algorithms that can enhance our daily clinical practice. In order to create effective algorithms using deep learning, massive ground truth data sets are required to train the algorithms. The production of such data sets can be time-consuming due to limitations of the image annotation interface (IAI). In order to rapidly produce large amounts of pathologist annotated data sets, first person shooter (FPS) video game principles, a type of game where speed and accuracy are keys to success, could be integrated into a pre-existing IAI. This modification has the potential to significantly improve the speed at which ground truth can be generated. Methods: Four pathology residents familiar with an unmodified version of a proprietary IAI were asked to report how many images they could annotate in one hour. The images were randomly chosen from a database of images of regions of interest (ROI). The IAI was then modified to integrate FPS principles of rounds-on-target (ROT), control ergonomics (CE), and alternate fire (AF). After the modifications, the same four pathology residents were tasked with annotating different sets of images and asked to report how many they could annotate in on hour. Results and Conclusion: Each of the participants showed great improvement with their annotation efficiency. Their improvements ranged from 140% to 200% by using the modified IAI. Integrating FPS principles of ROT, CE, and AF into an IAI is effective in improving the speed of ground truth generation.
| Predicting pN-Stage of Breast Cancer Patients with a Fully Automated Image Analysis System|| |
Hunter Jackson1, Richard Chen1, Yating Jing1, John Corbett1,2
1Research, Proscia Inc., Baltimore, MD, 2Department of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA. E-mail: [email protected]
Background: In order to prescribe accurate therapeutic plans for patients with breast cancer, patients must be evaluated by the TNM system. The TNM system provides a classification of cancer by anatomic disease extent; namely, the T category describes the primary tumor site, N describes the metastases in the regional lymph nodes, and M describes distant metastases. Our participation in the Camelyon17 challenge of the International Symposium on Biomedical Imaging (ISBI) tasked us with the challenge of investigating whether cancer has spread to the regional lymph nodes. Methods: Powered by deep convolutional neural networks and transfer learning approaches, we developed a fully automated system to detect regional lymph node metastases for breast cancer patients, thus providing a pN-stage for 100 unique patients. Using a novel stain normalization technique, we are able to eliminate inter-hospital variability in H and E staining and imaging of the slides, opening the door for global deployment of our algorithm using a cloud-based deployment mechanism. In this paper, we outline our data pre-processing techniques, the construction of our convolutional neural network, a novel approach to transfer learning, and the resulting automated system informed by this data. Results: Our fully automated system classifies a given patient's pN-stage with a kappa score of 0.766 indicating substantial agreement with expert pathologists. Conclusions: With a globally-scaled deployment of the algorithm via our cloud-based software, we are able to continue refining and validating our model as we approach clinical utility.
| Research Methodologies and Applications in WSI Analysis: A Perspective from an Analytic Microscopy Core Laboratory|| |
Joseph Johnson1, Agnieszka Kasprzak1, Tingan Chen1, Jonathan Nguyen1, Marilyn Bui1
1Analytic Microscopy Core Facility, H. Lee Moffitt Cancer Center, Tampa, Florida, USA. E-mail: [email protected]
The significance of digital pathology and image analysis in clinical oncology centers has become more evident in recent years. A centralized image analysis core plays an important role in bridging the expertise of clinicians, cancer biologists, and imaging scientists to address significant investigational questions. Here we will offer our experience with methodologies for digital WSI acquisition and image analysis; and demonstrate some challenges and limitations with these approaches. We will also present some examples of applying these methodologies which include: (1) dual stain immunohistochemical (IHC) analysis of human breast cancer xenographs, (2) serial section analysis of 9 biomarkers of metastatic melanoma tumors in lung sections, (3) Identification of Fox P3 positive lymphocytes in fluorescently labeled tissue microarray (TMA) slides, and (4) Second Harmonic Generation (SHG) imaging of collagen in mouse tumors. Finally, we will discuss some future directions in digital pathology.
| Trial of Consultation Portal to Support the Remote Diagnosis of Local Hospital|| |
Yukio Kashima1,4, Tomoo Ito2, Kenichi Yuma3, Hideki Okamura3, Masayuki Hayashi3, Han-seung Yoon4, Bungo Furusato4, Daisuke Niino4, Kazuhiro Tabata4, Takashi Koyama5, Junya Fukuoka4
1Department of Pathology, Awaji Medical Center, Sumoto, 2Department of Diagnostic Pathology, Kobe University Graduate School of Medicine, Kobe, 3Division of Healthcare, Philips Electronics Japan, Ltd., Tokyo, 4Department of Pathology, Nagasaki University Graduate School of Biomedical Sciences, Nagasaki, 5Administrative Office, Awaji Medical Center, Sumoto, Japan. E-mail: [email protected]
Background: Critical shortage of pathologists make it harder to render high-end diagnoses for specific clinical fields. Methods: Consultation portal named 'iReport-path' has been established in collaboration with Philips Japan to assist remote diagnosis of local hospital where one general pathologist who covers over 4000 cases annually. iReport-path is a remote pathology-report-management system to connect between supporting hospitals and local hospitals. Pathologists in supporting hospitals can manage all requests including ordering of additional staining in an integrated fashion and can finalize a case by uploading diagnostic reports to EMR of local hospital. If needed, additional consultation can be done by forwarding the case to other pathologists included in the list of the system. On the other side, Pathologists of the local hospital can send the case to the supporting hospital through LIS (Laboratory Information System). The final report of the case is transferred to LIS from iReport-path automatically. All connections were done through complete closed VPN. Results: Awaji Medical Center has requested 625 cases to NUH from Sep.,2016 to Mar., 2017 and 208 cases to KUH from Nov., 2016 to Mar., 2017 thorough iReport-path, respectively. All 833 cases were doublechecked and finalized by supporting hospitals. Additional expert advice was also mentioned in separate emails to the local pathologist when final diagnosis was changed from the original. All cases were processed within a reasonable turnaround time. Conclusion: Pathology consultation portal, iReport-path, is useful to support and improve the diagnosis of remote hospital and may overcome the shortage of pathologists.
| Digital Pathology: The Singularity Is Near|| |
Jong T. Kim1
1Pathology, Immunology and Laboratory Medicine, University of Florida Health, Gainesville, FL, USA. E-mail: [email protected]
Background: The technological singularity refers to a point in time when the artificial intelligence (AI) will surpass the human intelligence. Most technologists predict this will happen in our lifetime. It is also probable that continued advances in deep learning and AI will disrupt the digital pathology industry and ultimately how pathologists will practice. Therefore, understanding the landscape of how deep learning algorithms are being researched and applied to the digital pathology is imperative to the future practice of pathology. Methods: To systematically review 1) the current state of deep learning research in pathology, 2) the patent landscape in digital pathology, and 3) the AI-focused digital pathology startups Data Source: Deep learning research in pathology was searched using the PubMed search engine (www.ncbi.nlm.nih.gov/pubmed). Patents related to digital pathology were searched using the Google Patents search engine (www.google.com/patents). Google search (www.google.com), an online analytics service (www.cbinsights.com), and the startup incubators in the United States were searched for the AI-focused digital pathology startups. Conclusions: The PubMed literature search revealed that the deep learning research in pathology has increased significantly in the last decade with well over 200 publications so far. In addition, the number of patents granted, and patent applications filed in the field of digital pathology have also increased dramatically in the same time period. This surge in digital pathology research and development with deep learning has created a vast potential for artificial intelligence algorithms application in healthcare industry. The number of AI-focused startups in healthcare has increased from less than 10 in 2011 to more than 90 in 2016, many of which are in the digital pathology arena. The arrival of singularity in the digital pathology may by closer than we realize, and it is critical to the future of pathology that we start preparing our pathology trainees and department now for this eventual paradigm change.
| Alignment of Sequential Whole-Slide Images: A Deep Learning Approach|| |
Preston Law1, Beverly Faulkner-Jones2, Charles Law3
1Unaffiliated, 2Department of Pathology, Beth Israel Deaconess Medical Center, Boston, MA, 3Co-founder, Kitware, Inc, Clifton Park, NY, USA. E-mail: [email protected]
Background: Surgical pathology workflows involve sequential analysis of multiple slides and multiple stains. Traditional exam of sequential glass slides is skill-dependent and time-intensive. Digitization of slides and alignment of resulting whole slide images (WSIs) into a 3D stack has potential to improve diagnostic efficiency and accuracy. Fully automated alignment is hindered by tissue processing artifacts. Deep learning is an approach that can detect features across multiple sections. These features can then be used for automated alignment. Methods: Our deep network had a fully convolutional architecture, with 92x92 pixel receptive fields. A training set was generated by manual selection of 2978 glomeruli on 39 WSIs of sections stained with a variety of standard histochemical stains. The negative training images consisted of 8681 prior false-positive detections. After training, glomeruli were matched across sections and a homography was generated to align the stacks. We used gradient-descent to obtain the optimal transformation between glomerular sets. After regression, outlying false or missed detections were removed and the process was repeated. Results: Our final network achieved an 80% glomerular detection rate with minimal false positive errors. With additional annotated training images, this performance could be improved. Glomerular detection was accurate enough to align renal whole-slide image series. Conclusions: Deep learning is a powerful feature detection approach that is more flexible than traditional computer vision algorithms. With the addition of networks that detect other features, e.g. blood vessels, this technique may be generally applied to aligning sequential WSIs from clinical samples.
| Validation of Quantitative Digital Pathology Analyses|| |
Auranuch Lorsakul1, Joerg Bredno2, Jim Martin1, Shawn Wang2, Kien Nguyen1, Faith Ough3, June Clements3, Solange Romagnoli4
Background: Digital Pathology algorithms quantify the content of a whole slide or selected field of view (FOV) with respect to number of cells for one or more phenotypes. Especially when assessing the immune response to cancer, a count of immune cells like T-lymphocytes is primary read-out. Automated analyses require stringent verification to establish and assure the accuracy of cell counts. Methods: We propose to compare automatically generated cell counts to ground truth from expert pathologists in a framework that collects the following data: - Inter-observer agreement - Section-to-section agreement using aligned and registered FOVs - Algorithm-to-observer agreement Example studies are presented for the assessment of tumor cells and T-lymphocytes in a patient cohort with Stage II colorectal cancer. 4μm tissue sections were stained for CD3 (anti-CD3 2GV6) and CD8 (anti-CD8 SP238/57) on consecutive tissue sections. Two pathologists chose FOVs from a set of 119 slides stained with CD3 and 119 slides stained with CD8. On each slide, a pathologist chose three FOVs that represent tumor with high immune infiltrate, tumor with low immune infiltrate, and the invasive margin, respectively. The pathologists marked every T-cell in these FOVs. On 10 consecutive slide pairs, both pathologists provided the cell count in three FOVs to determine inter-observer variability. Results: 60 FOVs were available for the inter-observer study. The pathologists agreed with R2=0.957 for CD3 and R2=0.925 for the CD8 cell count, respectively. Total manual cell counts were 6,947 and 6,328 (ratio 0.911) for CD3 and 3,115 and 2,795 (ratio 0.897) for CD8, respectively. 714 manually counted FOVs were available for verification of the image analysis algorithm. Image analysis matched ground truth counts with R2=0.901 for CD3 and R2=0.943 for CD8, respectively. Total cell counts were 72,076 manual versus 66,179 automated (ratio 0.918) for CD3 and 34,133 manual versus 30,438 (ratio 0.891) for CD8, respectively. Conclusion: Digital pathology methods can provide reliable fully automated image analysis to assess density and location of cells IHC-stained tissue sections. A rigorous verification of these quantitative results is required to assess algorithm-to-observer agreement in relation to inter-observer agreement and section-to-section variability of cell counts.
| Label-Free Quantitative Breast Histopathology Using Spatial Light Interference Microscopy|| |
Hassaan Majeed1, Masanori Takabayashi2, Tan Nguyen3, Mikhail Kandel3, Zhen George Liu4, Andre Balla5, Gabriel Popescu3
Departments of 1Bioengineering and 3Electrical and Computer Engineering Beckman Institute of Advanced Science and Technology, University of Illinois at Urbana Champaign, 4Department of Pathology, University of Illinois at Urbana Champaign, 5Department of Pathology, University of Illinois at Chicago, Urbana, IL, USA, 2Department of Systems Design and Informatics, Kyushu Institute of Technology, Fukuoka, Japan. E-mail: [email protected]
Background: While Breast Cancer is a significant global health problem, the standard tissue evaluation method – manual investigation of a stained tissue biopsy by a pathologist – has shortcomings in its low throughput and inter-observer variability. Quantitative microscopy and digital pathology have the ability to address these shortcomings through quantitative and automated analysis. In this work we show that SLIM, a label-free quantitative imaging modality, is able to separate benign and malignant regions in breast tissue using quantitative metrics. Methods: Imaging contrast in SLIM is generated by measuring the optical path-length difference (OPD) across the tissue biopsy. We used the SLIM tissue scanner to image a tissue microarray comprising 68 cores (34 benign and 34 malignant). After extraction of OPD maps and assembling of the mosaic, the epithelial regions in each core were extracted using ImageJ. For each epithelial region, the mean scattering length l_s, the median curvature of the region C, the tissue disorder strength α and a texture metric G, extracted by using the Leung-Malik gradient filter bank, were used to separate benign and malignant epithelial regions. Results: Receiver Operating Characteristics (ROC) were used to measure the separation between the two classes. Results of three-fold validation showed an average AUC of 0.90. Conclusion: Our method can potentially provide pathologists with an objective and high-throughput tool for breast cancer screening and automated histopathology. This achievement can decrease the work-load for pathologists and provide an accurate method for early diagnosis leading to timely treatment.
| Histological Detection of High-Risk Benign Breast Lesions from Whole Slide Images|| |
Akif Burak Tosun1, Maurice Marx1, Luong Nguyen1, Nathan Ong1, D. Lansing Taylor1, S. Chakra Chennubhotla1, Olga Navolotskaia2, Gloria Carter2, Jeffrey L. Fine2
1Department of Computational and Systems Biology, University of Pittsburgh, 2Department of Pathology, Magee Womens Hospital of UPMC, Pittsburgh, PA, USA. E-mail: [email protected]
Accurate diagnosis of high-risk benign breast lesions is crucial in patient management since they are associated with an increased risk of invasive breast cancer development. Since it is not yet possible to identify the occult cancer patients without surgery, this limitation leads to retrospectively unnecessary surgeries. In this paper, we present a computational pathology pipeline for histological diagnosis of high-risk benign breast lesions from whole slide images (WSIs). Our pipeline includes WSI stain color normalization, ductal regions of interest (ROIs) segmentation, and cytological and architectural feature extraction to classify ductal ROIs into triaged high-risk benign lesions. We curated 93 WSIs of breast tissues containing high-risk benign lesions based on pathology reports and collected ground truth annotations from three different pathologists for the ductal ROIs segmented by our pipeline. Our method has comparable performance to a pool of expert pathologists.
| Transformative Role of Digital Pathology in Viral Neuropathogenesis Research and Development of Live Virus Vaccines against Neurotropic Viruses|| |
Olga A. Maximova1, Alexander G. Pletnev1
1Core Staff Scientist, Laboratory of Infectious Diseases, National Institute of Allergy and Infectious Diseases, Bethesda, MD, USA. E-mail: [email protected]
Background: There are many viruses capable of infecting the Central Nervous System (CNS) of humans. The viruses of our interest are tick-borne or mosquito-borne flaviviruses. Vaccines are available for some flaviviruses, with many vaccines being inactivated which makes them expensive and require extensive primary and boost vaccinations to maintain protection. Live virus vaccines are less expensive and can be more immunogenic, but the safety of such vaccines is always a concern. Only two successful live flavivirus vaccines are available – against yellow fever and Japanese encephalitis. Vaccines are still needed for most neuropathogenic flaviviruses. However, prior to clinical trials in humans, all newly developed vaccine candidates must show the evidence of low/acceptable neurovirulence potential in animal models (e.g., mice and nonhuman primates). Neurovirulence is evaluated in intracerebrally inoculated animals by routine microscopic analysis and semi-quantitative scoring of virus-mediated neuropathology. Methods: The number of animals and tissue slides representing major CNS regions of each animal are quite large. A routine microscopic analysis and semi-quantitative scoring is very labor-intensive and time-consuming process. To improve neurovirulence assessment of our vaccines, we adopted the digital pathology (DP). We created a large database of digital slides and improved CNS mapping and analysis using a DP workflow. We also developed the automated image analysis (AIA) of the cellular markers of neuroinflammation and neurodegeneration and validated its performance against semi-quantitative scoring. Results: DP workflow significantly improved time efficiency and reproducibility. Whole slide viewing and analysis at multi-magnification dramatically facilitated multi-brain analysis from mice models and neuroanatomical mapping throughout the entire CNS of NHPs. The correlation analysis revealed strong and significant positive correlations between the AIA and semi-quantitative scores. However, AIA is more objective and accurate. This enables quantitative assessment and differentiation of vaccine virus neurovirulence. Conclusions: The adaptation of DP to viral neuropathogenesis research and vaccine development programs will facilitate decision-making regarding vaccine safety and its advancement to clinical trials. Implementation of DP will depend on accessibility of hardware/software, as well as on experience and collaboration between pathologists in the scientific and regulatory settings. It is our hope that DP will markedly advance many fields of research including virology, immunology, neuroscience, and vaccinology. This research was supported by the Intramural Research Program of the NIAID, NIH.
| Implementation of Whole Slide Imaging as a Pathology Teaching Tool and for Institutional Tumor Boards: A resident's experience|| |
Ashish Mishra1, J Mark Tuthill1
1Pathology Informatics, Henry Ford Hospital, Detroit, MI, USA. E-mail: [email protected]
Background: This presentation will describe our experience implementing and utilization of whole slide imaging (WSI) as a teaching tool for the pathology residents in Henry Ford Hospital, Detroit as well as our initial efforts to use WSI at institutional tumor boards. Methods: Glass slides were scanned for practice over several weeks to determine basic operation, system performance and workflow processes. Experience quickly showed that the iScan HT could be used to improvement quality and efficiency of weekly unknown slide conference. A proposal was made and accepted to pilot this process. Initially there was a lot of reluctance and resistance from the group to use WSI. To increase interest and enthusiasm concordance studies were presented in the monthly journal club which compared WSI and glass slides. Results: In Oct 2016, unknown slide conference was presented using WSI. The reaction to the quality of the histopathology system usage was excellent: nuclear contours and nucleoli were clear; navigation easy; response time was excellent with no screen lag. The conference was well received. The residents and attending loved the new format. In November 2016, we started presenting cases on WSI in GYN tumor board. WSI, eliminated many problems went away in an instant rapid navigation to area(s) of interest, ease of switching between slides and ease of switching cases. Attendings were ecstatic after the tumor board. The resident was super confident and was able to explain everything in detail using WSI without fumbling. The Attending, clinicians and residents were enthused at the new format; some had no idea that this was even technically possible. Conclusion: Whole slide imaging is a useful tool for teaching and presentation purposes. It can be easily implemented and integrated into our day to day pathology practice and resident training. The reluctance to use WSI is initially high among pathologists, but enthusiasm increases once implemented into regular practice.
| Digital Microscopy Files of Archived Tumor and Derivative Bio-Specimens as Quality Control and Reference for Precision Medicine Standards. Essential Resources in Biorepository Operations and Future Biomedical Research|| |
Hector Monforte1, William Schleif1, Paris Volk1, Peter Steele1
1Anatomic Pathology and Pediatric Biorepository, Johns Hopkins All Children's Hospital, Saint Petersburg, FL, USA. E-mail: [email protected]
Background: Precision medicine grade and “Fit for Purpose” research tissue specimens and their derivatives destined for Biorepository archival necessitate an anonymized, practical and efficient manner of documenting their morphologic characteristics. High-resolution scanning of cyto/morphologic case material destined for RNA and DNA isolation and -80°C tissue archive is an essential requirement for meaningful and reproducible studies. Tumor heterogeneity, mandates a directed and thorough review of the features of the archived material or the source of genomic derivatives. Methods: Preparation of cell suspension slides and FFPE reference histologic sections of tumor specimens stabilized optimally (short ischemia time) in RNA stabilizing solutions, RPMI for DNA isolated from dissociated cells and representative tumor sections archived frozen in our biorepository can be routinely scanned with a Leica AT2 “Aperio” instrument. The resulting files exported de-identified within Leica eSlide Manager and displayed with Leica ImageScope©, available for reference to Biorepository operators and future investigators, documenting cyto/morphologic characteristics of tumor specimens from which RNA and DNA was extracted. Results: Virtual microscopy slides are more practical than actual slides (often unavailable) or captured images of case slide files, as they allow a thorough and full representation of case material, for review by potential interested researchers, to evaluate the characteristics of tissue, cells and material from which genomic material has been extracted. Conclusion: Representative tumor tissue should be submitted for cytologic and histologic preservation, and eventually for virtual microscope slide preparation as a reference of tumor-derived RNA, DNA and for all tissue fragments destined for archive in biorepository freezers.
| Japanese Guidance of Digital Pathology Diagnosis: Preparation for Guideline|| |
1Department of Pathology, Mita Hospital, International University of Health and Welfare, Minato-Ku, Tokyo, Japan. E-mail: [email protected]
Although digital pathology is spreading in Japan, most experienced pathologists feel nervous to make primary diagnosis by monitor images, not microscope. Japanese digital pathology developed on the base of static image telepathology which has been performed due to relative shortage of pathologists in rural area. Many big hospitals do not have full-time pathologist, and even the hospitals do have full-time pathologist, most of them are working alone. To backup these isolated pathologists, we are now working on government to cover the fee to remote double-check of pathology diagnosis using digital images. At least, this requires monitor diagnosis. So, we delivered “Guidance to make pathology diagnosis using digital pathology images” last year based on the know-how obtained by telepathology. Now we are trying to expand this relatively simple guidance to evidence-based-guideline.
| Deep Learning Nuclear Segmentation and Classification for Analysis of Lung Cytology Cell Blocks|| |
Paul G. O'Reilly1, Peter Bankhead1, Jim Diamond1, Peter W. Hamilton1
1Computational Pathology, Discovery and Education, Digital Pathology Solutions, Philips, Belfast, Northern Ireland. Email: [email protected]
Background: Cell-block preparation of lung cytology samples obtained by needle aspiration is beneficial in providing diagnostic information in lung cancer and is used to provide material for molecular testing of mutations. One of the challenges in carrying out molecular testing is ensuring sufficient tumor nuclear material in the sample macrodissected for downstream analysis. This is typically carried out by a trained pathologist, but there is evidence for large variability in the estimation of tumor cell percentage by pathologists. Automated methods of identifying tumor-rich regions for macrodissection offer benefits in reduced variability in lung histology specimens. The objective of this work is to investigate a novel Deep Learning method of identifying tumor cell nuclei in whole slide images of lung cytology cell block specimens. Deep Learning is a data-driven machine learning approach to image classification and segmentation which has shown considerable utility in image processing, including increasingly in Computational Pathology applications. Methods: We propose a detailed annotation process for obtaining ground truth data from cell block images, which are used to train a Convolutional Neural Network and validated against an unseen set of annotations obtained from the same set of slide images. Results: Pixel-level accuracy metric allows performance quantification, ranging from 63.8%-99.7% (mean = 90.0%) for the validation set. Conclusions: The algorithm, applied to whole slide images, provides a slide-level segmentation of tumor and non-tumor nuclei. Using simple peak detection, it can provide an estimate of tumor nuclei/non-tumor nuclei ratio across the slide.
| Use of Digital Pathology to Improve Diagnostic Accuracy in Breast Cancer Reporting to a Qualified Clinical Data Registry Measure Collection by CMS|| |
Sandra Martins1, Patricia Goede1, Zaibo Li2, Anil Parwani2
1Research and Development–Informatics, XIFIN, Inc., San Diego, CA, 2Department of Pathology, Wexner Medical Center, The Ohio State University, Columbus, OH, USA. E-mail: [email protected]
Background: Qualified Clinical Data Registries (QCDR) capture quality measures as part of PQRS. A QCDR is a CMS-approved entity that collects medical and/or clinical data for patient and disease tracking to foster improvement in the quality of care provided to patients. Accurate reporting to a QCDR is dependent on review of digital pathology as a critical component to pathology diagnosis and the ability to report accurate CPT and ICD codes that can drive reporting to a QCDR. Methods: IRB approval was obtained for a retrospective review of 60 randomly selected breast pathology reports that were diagnosed at Ohio State University. All cases, including review of slides by digital pathology was included with the pathology report for an accurate and appropriately coded diagnosis. Breast histologic type pT, pN was included to meet the CMS requirement for quality reporting. Results: The OSU study consisted of 60 Breast randomly selected pathology reports and mix of race and positive and negative results. Lack of CPT codes for ER/PR negative results along with proper stage reporting presents challenges in the QCDR. 36 cases were ER/PR positive while the remaining cases were mixed. Conclusions: Pathologists remuneration will be calculated on the quality of information submitted to CMS. Operations may be hindered to accurately submit data to a QCDR. As the quality scores will be made public, reputations may be negatively impacted. The use of digital pathology improves quality as part of the overall clinical information system and is a key component to accurate reporting to CMS.
| Pattern Recognition and Quantification of Hepatic Fibrosis in NASH Preclinical Models Using Deep-Learning Based Image Analysis|| |
Elton Rexhepaj1,2, Nathalie Degallaix1,2, Benoit Noel1,2, Carole Belanger1,2, Robert Walczak1,2, Sophie Megnien1,2, Bart Staels1,2, Dean Hum1,2, John Brozek1,2
1Data Scientist, GENFIT, Loos, 2Institute Pasteur de Lille, University Lille 2, Lille, France. E-mail: [email protected]
Background: Digital pathology has been transformed in recent years by the development of whole-slide imaging systems. Although the assessment of advanced liver fibrosis is generally made with good accuracy, the quantification of early stages of fibrosis (e.g. perisinusoidal fibrosis) is challenging with moderated inter-observer agreement (κ=0.5) (2). Deep-learning pattern recognition algorithms have recently shown to be able to capture accurately both micro and macro tissue patterns, mimicking the diagnostic workflow of the human pathologist (3). The aim of this work was to establish a fully integrated and automated pipeline for the assessment of NASH fibrosis staging using liver picro-sirius sections. Methods: Digital glass slides were processed to extract high magnification representations (4x, 2μm). A deep-learning model was constructed based on 683 fields of view (FOV, 512x512 pixels) representing the main histological patterns of NASH fibrosis (1: Negative-collagen, 2: Perisinusoidal-collagen, 3: Periportal-collagen, 4:Vascular-collagen, 5:Bridge-collagen) divided into a training set (N=511) and a validation set (N=172). Results: We show how a deep-learning framework can be used to automatically recognize: (a) collagen-negative regions (88% accuracy across 155 FOVs), (b) vascular regions (86% accuracy across 155 FOVs), (c) collagen-positive portal-triad (88% accuracy across 99 FOVs), (d) collagen-positive sinusoids (79% accuracy across 119 FOVs), (e) bridging fibrosis (99% accuracy across 155 FOVs). Conclusion: We made use of a deep-learning algorithm to detect fibrosis; more complex algorithms to evaluate ballooning and lobular inflammation stages are still under development and should help in assessment of NASH disease activity and the overall liver histology in the near future.
| The Artificial Pathologist: Deep Learning Automatically Identifies Pathologically Significant Features|| |
Drew Linsley1, Thomas Serre1, Anthony Magliocco2, Daryoush Saeed-Vafa2
1Department of Cognitive, Linguistic, and Psychological Sciences, Brown University, Brown Institute for Brain Science, Providence, RI, 2Department of Anatomic Pathology, H. Lee Moffitt Cancer Center and Research Institute, Tampa, FL, USA. E-mail: [email protected]
Background: Computer vision-assisted histopathological analysis has great potential for automating pathological diagnoses and improving prognoses. These tools often draw upon domain expertise to target biologically significant morphology. In contrast, state-of-the-art artificial vision typically leverages deep convolutional networks (DCN), which automatically learn to detect optimal features for the task at hand. Here we measure DCN performance and identify the morphological features they use for diagnosing metastatic breast cancer in lymph nodes (LN) on whole slide images (WSI). Methods: A DCN was trained to distinguish benign from malignant image patches extracted from 260 annotated, hematoxylin and eosin stained, LN WSIs. Performance was evaluated as malignancy detection accuracy on WSI image patches held-out from training. Morphological features used by the DCN were identified with layerwise-relevance propagation (LRP). Results: The DCN was 98% accurate in detecting malignancy after approximately 50 epochs of full-dataset training. LRP identified the features it used, which highlighted nuclear size and cherry-red nucleoli. Conclusions: Joining others, we find strong evidence that DCNs are well-suited for screening LNs for malignancy. However, it has been unclear if DCNs use similar morphology as pathologists for this task. We demonstrate that DCNs automatically learn pathologically significant features during training to screen LNs for malignancy. To our knowledge, this is the first study systematically examining how DCNs achieve their performance in WSI analysis. By opening up the DCN black-box we find that DCNs use qualitatively similar features to pathologists for diagnosis, validating their tissue screening performance and demonstrating their transformative potential for computational pathology.
| Qualitopix: Automated Quality Assessment of the HER2 Receptor|| |
Stine Harder1, Astrid Ottosen1, Andreas Schønau1, Keith Miller2
1A/S, Visiopharm, Hørsholm, Denmark, 2Director, United Kingdom National External Quality Assessment Service, London, United Kingdom. E-mail: [email protected]
Background: Immunohistochemistry is an important tool in patient diagnostics, and external quality assessment (EQA) of immunoassays is essential to obtain optimal and comparable results. EQA is usually carried out by 4 independent expert assessors using a multiheaded microscope. Occasionally there are discrepancies, and this can be due to the volume of slides that are being assessed. Digital image analysis (DIA) may support EQA by extracting useful information to provide objective and standardized results. Our method is one step in the direction of using DIA for automatic quality assessment. Methods: The data consisted of four HER2 cell lines with various HER2 expression used in the UKNeqas RUN 111. The slides were immunostained by 54 laboratories for HER2 and assessed by pathologists at UKNeqas as poor (7), borderline (6), good (16), and optimal (25), with poor being insufficient and the remaining categories sufficient. DIA was furthermore used to analyze the cell lines with respect to five different characteristics; amount of membrane staining, intensity of membrane staining, intensity of cytoplasm, intensity of nucleus, and number of small fragments/artefacts. The DIA results assessed the laboratories as insufficient or sufficient using an individual threshold value for each core and characteristic. Results: Agreement between DIA and pathologist assessment was found for 97% of the laboratories assessed as poor or optimal by the pathologist and for 77% of the laboratories including the intermediate categories, borderline and good. The intermediate categories also showed the largest inter-pathologist variance, why it is not surprising that the DIA algorithm has the lowest agreement for those assessments. Furthermore, a dataset including more insufficient cases would be needed for a better training of the algorithm. Conclusion: Our image analysis method for quantifying staining sufficiency using HER2 cell lines shows a promising step in the direction of using DIA for quality assessment.
| A 3-year Report on Whole Slide Imaging Program for Clinical Slide Archiving|| |
Matthew Suriawinata1, Heather Warren1, Gerald Jackman1, Laura Gordon1
1Department of Pathology and Laboratory Medicine, Dartmouth-Hitchcock Medical Center, Lebanon, New Hampshire, USA. Email: [email protected]
Background: Although whole slide imaging has been widely adopted, its implementation in a clinical pathology practice remains challenging. The Department of Pathology at Dartmouth-Hitchcock Medical Center (DHMC) has successfully implemented a routine whole slide scanning (WSS) for archiving in the past 3 years. Methods: The Department of Pathology at DHMC has been performing WSS program for archiving since 2014, initially using two Leica SCN400 scanners, which were replaced by two Leica AT2's and one CS2, employing the same workflow. One to four pertinent cytopathology, dermatopathology, surgical pathology and consultation slides, are selected and documented in pathology report. These slides are submitted for scanning, immediately after the case was signed out. Whole slide images (WSI) are reviewed for quality and uploaded to hospital data center. Scanned slides are marked and filed in the slide storage room. Reasons for retrieval of scanned slides are recorded. Results: The total numbers of slides scanned were 29,066 (2014), 43,214 (2015), 48,090 (2016), representing 19%, 42%, and 46% of clinical cases respectively. The average turnaround time for routine and priority scanning were 24 and 0.5 hours. The service requires 0.75 FTE support and is managed by histotechnologists due to familiarity on histology quality control. All pathologists now rely on WSI for review and secondary diagnosis. Storage of WSI files at hospital data center ensures security and reliability. Conclusion: WSS can be integrated to a busy pathology service, with a well-designed workflow and reliable histotechnologist support. The program familiarizes pathologists with WSI in preparation for primary diagnosis in the future.
| Computational Discovery of Tissue Morphology Biomarker for Very Long-Term Survivors with Pancreatic Ductal Adenocarcinoma|| |
Jacob S. Sarnecki1,2, Laura D. Wood3,4, Ralph H. Hruban3,4, Anirban Maitra5,6, Denis Wirtz1,2,4, Pei-Hsun Wu1,2
1Johns Hopkins Physical Sciences-Oncology Center, The Johns Hopkins University, 2Department of Chemical and Biomolecular Engineering, The Johns Hopkins University, 3Department of Pathology, The Sol Goldman Pancreatic Cancer Research Center, Johns Hopkins University School of Medicine, 4Department of Oncology, Johns Hopkins University School of Medicine, Baltimore, Maryland, 5Department of Translational Molecular Pathology, The University of Texas MD Anderson Cancer Center, 6Sheikh Ahmed Bin Zayed Al Nahyan Center for Pancreatic Cancer Research, The University of Texas MD Anderson Cancer Center, Houston, Texas, USA. E-mail: [email protected]
Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest forms of cancer, with an average 5-year survival rate of only 8%. Within PDAC patients, however, there is a small subset of patients that are able to survive for more than 10 years. Deciphering underlying reasons behinds prolonged survival potentially provide new opportunity to treat PDAC and yet, no genomic, transcriptomic, proteomic, or clinical signatures have been found to robustly separate this subset of patients. Digital pathology in combination with machine learning provides the opportunities to computational search the tissue morphology patterns in associating with disease outcome. In this work, we developed the computational framework to analyze whole-slide images (WSI) of PDAC patient tissues and identified the tissue morphology signatures for very long term surviving patients. Our result indicates that low tissue morphology heterogeneity is significantly linked to better survival. Furthermore, regionally extra-tumoral tissue encodes prognostic information for patient survival. Based on information from both tissue morphology as well as tissue heterogeneity in tumor and its adjacent area we established a machine learning model with an AUC of 0.94. In sum, our study demonstrates a pathway to accelerate the discovery of undetermined tissue morphology in association with pathogenesis states and patient outcome for prognosis and diagnosis by utilizing computational approach with digital pathology.