Arpit Aggarwal
BME PhD Defense Presentation

Date: 2025-06-10
Time: 2:30pm - 4:30pm
Location / Meeting Link: Emory HSRB II 6th Floor (Virtual via zoom link: https://emory.zoom.us/j/4346780350?omn=98065827609)

Committee Members:
Anant Madabhushi, PhD (Advisor); Saurabh Sinha, PhD; Ahmet F. Coskun, PhD; Susan Modesitt, PhD; T. Rinda Soong, PhD

Title: Artificial Intelligence-Based Phenotyping of the Tumor Microenvironment in Hematoxylin and Eosin-Stained Images of Solid Tumors

Abstract:
The tumor microenvironment (TME) of solid tumors plays a crucial role in cancer progression and treatment response. Pathologists traditionally rely on Hematoxylin and Eosin-stained (H&E) images for tissue examination. However, conventional assessments of H&E images are qualitative, leading to interobserver variability and limited reproducibility. To address these challenges, this work develops artificial intelligence (AI)-based methodologies to quantitatively analyze both visible and non-visible components of the TME in digitized H&E images. Visible components, such as tumor-infiltrating lymphocytes (TILs), exhibit distinct morphological characteristics that make them identifiable in H&E images. In contrast, non-visible components, such as M2-subtype tumor-associated macrophages (M2-TAMs), lack clear morphological characteristics in H&E images and require additional staining for detection. The first part of this work focuses on quantitatively analyzing visible TME components and their association with treatment response and outcomes in solid tumors. We introduce CollaTIL, an AI-based method that characterizes the relationship between TILs and collagen, providing prognostic insights in gynecologic cancers. We also examine the immune and collagen changes in high-grade serous ovarian carcinoma before and after neoadjuvant chemotherapy, identifying structural alterations in the TME linked to treatment response. Additionally, we present an AI-based method for assessing tumor-stroma proportion in H&E images, demonstrating its superiority over biomarkers like KELIM in predicting treatment response and outcomes in epithelial ovarian cancer. The second part explores AI-based methodologies for quantitatively analyzing non-visible TME components, such as M2-TAMs, and their association with outcomes in solid tumors. We introduce VISTA, an AI-based method that translates H&E images into virtual immunohistochemistry-stained images, enabling the identification of M2-TAMs without additional staining. By integrating AI-based analysis of the TME in H&E images with treatment response and outcomes in solid tumors, this work advances computational pathology and contributes to the development of more precise prognostic tools.