Project 3: AI Digital Pathology Tools to Address Disparities in Lymphoma Diagnosis and Classification for Underserved Communities

University Of Tx Md Anderson Can Ctr, Houston TX

Investigators

Abstract

PROJECT 3_ AI DIGITAL PATHOLOGY TOOLS TO ADDRESS DISPARITIES IN LYMPHOMA DIAGNOSIS AND CLASSIFICATION FOR UNDERSERVED COMMUNITIES : ABSTRACT Given the complexity of lymphoma diagnosis, community pathology practices are at a distinct disadvantage in diagnosing and classifying lymphomas due to lack of onsite subspecialty hematopathology expertise and access to advanced immunophenotyping and molecular testing. Patients who receive a pathologic diagnosis of lymphoma at a community-based practice are also more likely to experience delays and have higher costs for diagnosis than patients diagnosed at an academic medical center. Their lymphoma is misdiagnosed 20% of the time, often with treatment-changing implications. These challenges disproportionately impact populations that have worse outcomes and rural groups who receive care at community-based hospitals and centers. Digital pathology presents new opportunities for addressing diagnostic issues in these settings and has emerged as one of the most successful clinical applications for artificial intelligence (AI), enabling the use of AI tools to assist generalists with tasks like cancer detection and grading. These approaches have mainly been applied in solid tumors; hematopathology applications, particularly those for lymphomas, have received less attention due to a paucity of WSI datasets with clinical annotation, the complexity of diagnostic systems, and their reliance on extensive immunohistochemical (IHC) staining. Our group and others have developed AI models that predict biomarkers from WSIs of hematoxylin and eosin (H&E)-stained slides, which can provide an inexpensive screen to rule out or replace IHC and molecular testing for biomarkers. However, skewed classification remains a challenging problem for AI tools, due in part to lack of representation of some groups in datasets. Our research team is uniquely equipped to address these challenges, given our prior work in lymphoma classification, digital pathology and AI tool development; experience integrating community and academic hematopathology practice; and access to the Lymphoma Epidemiology of Outcomes (LEO) cohorts . We will use LEO to develop and validate a diagnostic lymphoma classifier that provides highly accurate diagnosis from WSI of a simple H&E stain that can be performed in any pathology lab, with dedicated bias mitigation techniques to enhance diagnostic accuracy across populations. We will also investigate how integrating spatially resolved molecular data into a machine-learning model can create more powerful H&E-based models that predict prognostic biomarkers of the lymphoma microenvironment. These models will be built specifically to meet the needs of community-based practices and implemented in community healthcare systems, with the ultimate goal of improving clinical outcomes of the poor-risk and rural groups that they serve.