To view the video recordings of these sessions, go to our YouTube playlist for this workshop.

Day 1 (Tuesday, January 26, 2021)


Keynote Talk #1

Speaker: Ulysses Balis, University of Michigan
Title: Supporting the NIDDK Kidney Precision Medicine Project (KPMP): Standing up the U-M Pathology AI / Data Visualization Center and Core Lab – Five years in retrospect [Slides]
Abstract: The NIDDK Kidney Precision Medicine Project (U2C mechanism) is a multi-institutional consortial effort to generate further biological understanding of acute kidney injury and chronic kidney disease, through a novel discovery model that includes actual study participants (who serve concurrent roles as being both study subjects and patients at the same time), with the goal of elucidating new renal cell types, regulatory and disease mechanisms, and therapeutic targets in the process of developing a multi-axial atlas of combined -omics, single-cell, and histology data. The Data Visualization Center (DVC) at the University of Michigan (UM) plays a central role in the creation of the needed tools for this effort, with one thrust being the development of a Whole Slide Image computational discovery suite, known as VIPR Studio. This presentation will provide an overview of this tool suite, the UM Pathology Imaging Core Lab, and how these efforts are shaping the KPMP effort.

Invited Talk #1

Speaker: Michael Becich, University of Pittsburgh
Title: Computational Pathology and Improving Predictive Analytics [Slides]
Abstract: With the FDA approval of whole-slide imaging (WSI) and digital pathology workflows as part of Pathology practice, there is a significant gap in infrastructure for supporting translational research in cancer centers. Pathology Informatics is key to supporting deep interrogation of WSI and allowing for multidimensional studies combining multiplexed immunohistochemical marker studies of the tumor microenvironment with the rich genomic data provided by single-cell genomic studies. Modern tissue-banking protocols coupled with diagnostic workflows in surgical pathology of patient’s tumor samples are key to supporting both computational pathology and predictive analytics. We will discuss an approach that incorporates multiple sources of pathology data and clinically actionable knowledge to improve Pathology decision-making via predictive analytics fueled by advanced machine learning and causal discovery.

Invited Talk #2

Speaker: Anant Madabhushi, Case Western Reserve University
Title: Digital Pathomics – An Alternative to Deep Learning to Prognosticating Disease Outcome [Slides]
Abstract: With the advent of digital pathology, there is an opportunity to develop computerized image analysis methods to not just detect and diagnose disease from histopathology tissue sections, but to also attempt to predict risk of recurrence, disease aggressiveness, and long-term survival. At the Center for Computational Imaging and Personalized Diagnostics, our team has been developing a suite of image-processing and computer-vision tools, specifically designed to predict disease progression and response to therapy via the extraction and analysis of image-based “histological biomarkers” derived from digitized tissue-biopsy specimens. These tools would serve as an attractive alternative to molecular-based assays, which attempt to perform the same predictions. The fundamental hypotheses underlying our work are that: (1) the genomic expressions detected by molecular assays manifest as unique phenotypic alterations (i.e., histological biomarkers) visible in the tissue, (2) these histological biomarkers contain information necessary to predict disease progression and response to therapy, and (3) sophisticated computer-vision algorithms are integral to the successful identification and extraction of these biomarkers. We have developed and applied these prognostic tools in the context of several different disease domains, including ER+ breast cancer, prostate cancer, Her2+ breast cancer, ovarian cancer, and more recently, medulloblastomas. For the purposes of this talk, I will focus on our work in breast, prostate, rectal, oropharyngeal, and lung cancer.

Invited Talk #3

Speakers: Jason Hipp and Khan Baykaner, AstraZeneca
Title: How Digital Pathology Will Transform Drug Development and Cancer Diagnosis
Abstract: Digital pathology, with the advent of modern deep-learning techniques, promises to revolutionize drug development and cancer diagnosis. Traditional pathology requires highly trained individuals to make judgements based on time-constrained analyses of extremely rich and complex data. AI and automation are ideally suited to such tasks; whereas it is expensive and difficult to scale-up the population of highly trained pathologists, it is cheap and easy to scale up digital computation. Modern deep-learning techniques have shown amazing advances across domains such as natural image processing, natural language processing (NLP), and speech recognition. In this talk, Drs. Jason Hipp and Khan Baykaner will explain how these techniques can be applied to pathology and drug development, and the way in which this unlocks previously unavailable analyses.

Invited Talk #4

Speaker: Jeffery Goldstein, Northwestern University
Title: Machine learning for placental pathology: why, how, and what now? [Slides]
Abstract: The placenta is the first organ to form in human development and acts as the lungs, gut, and kidney in fetal life. Most adverse pregnancy outcomes are associated with abnormalities in the placenta—abnormalities that may have lifelong consequences for the mother or infant. Yet, <20% of placentas delivered in the USA receive a pathologic examination. Among pathologists, there is a wide range of capability and high interobserver variability. Machine learning and digital pathology offer the possibility of making placenta examination universally available and increasing consistency. As a demonstration of the potential utility of machine learning in placental pathology, we trained a model to estimate gestational age from the microscopic appearance of placental villi. Gestational age is the single most important variable in neonatal survival and abnormalities in the appearance of villi—either appearing from a later (accelerated maturation) or earlier (delayed maturation) gestational age—are associated with the pregnancy complications of pre-eclampsia and diabetes, among others. We scanned slides from 154 patients representing a range of gestational ages from 24 weeks, the threshold of viability, to 42 weeks, after the expected time of delivery (37–40 weeks). To simulate gestalt formation by pathologists, we trained a model using 16 high-power fields at a time, scattered across a single slide. The model contains an attention subnetwork that weights each high-power field before generating an average. Using the attention and aggregation, we showed a high accuracy. We tested the utility of this model in estimating gestational age from unannotated whole-slide images, with high accuracy. This work shows the utility of machine learning in placental pathology. One use case could be providing an estimated gestational that pathologists can use to improve interrater reliability in their diagnoses of appropriate/delayed/accelerated maturation. This use case highlights key challenges, including interinstitutional slide variability, slide artifacts, differences in scanners or cameras, pathologist disagreements about regions of interest, and interpretability. Robust solutions to these problems will increase the utility of AI and the digital pathology field.

Day 2 (Wednesday, January 27, 2021)


Keynote Talk #2

Speaker: Raghu Machiraju, The Ohio State University
Title: Highlighting Challenges for Machine Learning in the Pathology Clinic through Specific Use Cases [Slides]
Abstract: In this talk, I will provide case studies drawn from both well-characterized and not-so-well-studied cancers of the breast, the thyroid, the prostate, and sarcomas in general. Each case study will address the reasons as to why deploying machine learning (ML) is hard, given the intrinsic complexity and a well-acknowledged paucity of training data. In the same vein, through appropriate demonstrations, I will illustrate the difficulties in gathering training data from whole-slide images. Further, these case studies will be examined for their inability to make robust and interpretable predictions.

Invited Talk #5

Speaker: Eric Fosler-Lussier, The Ohio State University
Title: Understanding how to understand teammates [Slides]
Abstract: Language is perhaps our most fundamental communication medium, and creating natural interactions within human-machine teaming can only enhance the collaborative efforts of all partners in the team. However, language is often imprecise, relying on inference and context to resolve ambiguities, which can be difficult in the face of a sparse data landscape. This talk has two parts: the first part will be a high-level discussion of the challenges that come with situated, teamed-language understanding and generation in human-machine teaming systems. The second part discusses some of our work in mitigating issues of uncertainty, error, and sparseness in understanding for spoken dialogue systems, drawing on examples from a Virtual Patient system used to train medical students.

Invited Talk #6

Speaker: Srinivasan Parthasarathy, The Ohio State University
Title: Stochastic Flow Clustering: Consolidation, Renewed Bearing and Applications to Image Segmentation [Slides]
Abstract: Since its introduction in the late nineties, the idea of Markov Clustering, a graph-clustering approach based on the principle of simulating stochastic flows (random walks) has seen wide use. In the first part of this talk, I will review this basic idea and then describe several principled enhancements to this approach that, in turn, improve the quality (via regularization, and the accommodation of overlapped clustering) and speed (via sparsification, and a multilevel mechanism) of such stochastic flow algorithms so that they can be deployed on big data problems. I will subsequently discuss ongoing efforts on leveraging these ideas for human-in-the-loop image analysis.

Invited Talk #7

Speaker: José Otero, The Ohio State University
Title: Providing molecular insight for resource-strained patients with machine learning-based workflows [Slides]
Abstract: The advent of molecular testing has improved medical decision-making by increasing the precision in which neuropathological diagnoses predict patient outcomes. However, many molecular tests, even those recommended by international clinical guidelines, are often not reimbursed by health insurance providers. To save costs, tests that increase the pretest probability of molecular testing will prevent waste, lower costs, and accelerate the delivery of information needed for treatment-plan generation. As a test case, we evaluated the diagnostic workflow for oligodendroglioma, a well-studied WHO grade II neoplasm with a prognosis relatively favorable astrocytoma. The distinction between these entities is based on a chromosomal deletion in chromosomes 1p and 19q. This test can be performed by fluorescent in situ hybridization (FISH), a reimbursable test with a false-positive rate of ~5%, or by chromosomal microarray, a test considered a gold standard but for which reimbursement is not possible. Our clinical decision-making challenge was to determine when we should order the nonreimbursable test in the setting of a positive FISH result. We applied Bayesian probability, information theory, and dimensionality reduction techniques to identify and compare the utility of distinct assays in improving medical decision-making.

Invited Talk #8

Speaker: Hamid Tizhoosh, University of Waterloo, Canada
Title: BERT, Transformers, NLP, and Pathology Reports [Slides]
Abstract: The advent of molecular testing has improved medical decision-making by increasing the precision in which neuropathological diagnoses predict patient outcomes. However, many molecular tests, even those recommended by international clinical guidelines, are often not reimbursed by health insurance providers. To save costs, tests that increase the pretest probability of molecular testing will prevent waste, lower costs, and accelerate the delivery of information needed for treatment-plan generation. As a test case, we evaluated the diagnostic workflow for oligodendroglioma, a well-studied WHO grade II neoplasm with a prognosis relatively favorable astrocytoma. The distinction between these entities is based on a chromosomal deletion in chromosomes 1p and 19q. This test can be performed by fluorescent in situ hybridization (FISH), a reimbursable test with a false positive rate of ~5%, or by chromosomal microarray, a test considered a gold standard but for which reimbursement is not possible. Our clinical decision-making challenge was to determine when we should order the nonreimbursable test in the setting of a positive FISH result. We applied Bayesian probability, information theory, and dimensionality reduction techniques to identify and compare the utility of distinct assays in improving medical decision-making.

Day 3 (Thursday, January 28, 2021)


Keynote Talk #3

Speaker: Lee Cooper, Northwestern University
Title: Charting a future course for computational pathology [Slides]
Abstract: There have been significant advances in both machine learning (ML) and digital pathology over the past decade, and the performance of ML in pathology applications suggests clinical utility in the not-so-distant future. In this talk, I will discuss recent history of ML for pathology imaging, describe limitations of current approaches based on deep learning, and discuss challenges that need to be overcome to increase clinical translation. These challenges include scaling data labeling and annotation, creating realistic validation datasets that can reveal weaknesses in ML models, and the creation of shared data resources.

Invited Talk #9

Speaker: Chakra Chennubhotla, University of Pittsburgh
Title: Spatial Analytics with Explainable AI for Anatomic Pathology [Slides]
Abstract: Pathologists are adopting whole-slide images (WSIs) for diagnosis, thanks to recent FDA approval of WSI systems as class II medical devices. In response to new market forces and recent technology advances outside of pathology, a new field of computational pathology has emerged that applies artificial intelligence (AI) and machine-learning algorithms to WSIs. Computational pathology has great potential for augmenting pathologists’ accuracy and efficiency, but there are important concerns regarding trust of AI due to the opaque, black-box nature of most AI algorithms. In addition, there is a lack of consensus on how pathologists should incorporate computational pathology systems into their workflow. To address these concerns, building computational pathology systems with explainable AI (xAI) mechanisms is a powerful and transparent alternative to black-box AI models. xAI can reveal underlying causes for its decisions; this is intended to promote safety and reliability of AI for critical tasks such as pathology diagnosis. This talk outlines xAI-enabled applications in anatomic pathology workflow that improve efficiency and accuracy of the practice. In addition, we describe HistoMapr-Breast, an initial xAI-enabled software application for breast core biopsies. HistoMapr-Breast automatically previews breast core WSIs and recognizes the regions of interest to rapidly present the key diagnostic areas in an interactive and explainable manner. We anticipate xAI will ultimately serve pathologists as an interactive computational guide for computer-assisted primary diagnosis.

Invited Talk #10

Speaker: Mike Montalto, PathAI
Title: AI-based pathology in clinical stage bio-pharmaceutical drug development [Slides]
Abstract: Pathology-based biomarkers play an important role along the continuum of early- and late-phase biopharmaceutical drug development, including indication selection, pharmacodynamics, and patient selection and stratification. In the era of precision medicine, traditional microscope-based pathology may fall short of the needed precision, reproducibility, and accuracy that are needed to enhance the technical and regulatory success of new drug applications. This talk will explore the application of machine learning and artificial intelligence (AI) in pathology to the clinical drug-development paradigm. Real-life examples will be reviewed in oncology and non-oncology therapeutic areas where AI-based pathology provides superior performance over traditional microscopic methods. Regulatory considerations of using AI-based pathology in drug development will also be discussed.

Invited Talk #11

Speaker: Mohamed Salama, Mayo Clinic
Title: AI Promise to the Practice of Hematopathology [Slides]
Abstract: Hematopathologists are increasingly using digital-imaging tools for a wide spectrum of practice settings. However, digital imaging associated with artificial-intelligence applications for effective learning and diagnosis rendering are not yet routinely incorporated in practice. We will share our experience in utilizing digital tools and will demonstrate methods and applications of digital imaging along with augmented human intelligence to effectively improve the workflow in the practice of hematopathology. We will cover the essential elements as well as the pitfalls, advantages, challenges, and opportunities in utilization of digital tools in practice.

Invited Talk #12

Speaker: Hari Subramoni, The Ohio State University
Title: High-Performance Deep Learning with Large Pathology WSI Images [Slides]
Abstract: This talk will focus on high-performance computing (HPC) technologies, techniques, and solutions to accelerate various operations associated with computational pathology. In the distributed Machine Learning (ML) training regime, data parallelism has become an established paradigm to train ML models that fit inside GPU memory on large-scale HPC systems. However, model parallelism is required to train very large whole-slide pathology images and associated models that will not fit in the memory available on modern GPU platforms. We deal with emerging requirements brought forward by such very large models and whole-slide images (WSIs) being trained using high-resolution images common in digital pathology. To address these, we propose, design, and implement GEMS, a GPU-Enabled Memory-Aware Model-Parallelism System. We present several design schemes that offer excellent speedups over state-of-the-art systems like Mesh-TensorFlow and FlexFlow. Our designs enable the training of large ML models on large patches of histopathology images and decrease the overall training time. The lower training time allows the deep-learning researcher to train bigger and better models, thereby increasing existing designs’ accuracy. We show the potential of our proposed designs by training a custom ResNet-110-v2 on image tiles of size 1,024 × 1,024 and reducing the training time from 1 day to 28 minutes for the real-world histopathology WSI of 100,000 × 100,000 pixels.