Exploring Explainable AI : Differential Diagnosis of Benign Breast Lesions

Abstract
The differential diagnosis of breast lesions is challenging and remains a critical component of breast cancer screening, presenting even for experienced pathologists a more difficult classification problem than the binary detection task of cancer-vs-not-cancer. Within the context of real pathologist work, it includes careful assessment for alterations in breast ducts as they make diagnostic decisions. Particularly, they continually observe tissue patterns and make decisions supported by the morphology. In doing so, they might look at an entire duct or at patterns occurring within selective portions of the duct striving to generate mental associations with similar ducts and/ or parts (prototypical) that they previously encountered during pathology training or subsequent clinical practice. Currently, there are no methods which provide directions for a computational understanding of the structural changes in the breast tissue triggered by atypia and other malignancies. To address this, we reframe the computational diagnosis of breast lesions as a problem of prototype recognition on the basis that pathologists mentally relate current histological patterns to previously encountered histological patterns (prototypes) during their routine diagnostic work. We address this by analytically modeling a visual pattern dictionary that traditionally defines the standards on tumor classification/ nomenclature for pathologists worldwide. A broader aim of my thesis is to build an end-to-end computational pathology framework intended to resonate with the visual diagnostic thinking of the pathologists.
About the Speaker :
Akash Parvatikar is a PhD candidate in Computational Biology at the Joint Carnegie-Mellon and University of Pittsburgh School of Medicine specializing in biomedical imaging informatics and computational and systems pathology. He is interested in investigating the intrinsic characteristics of biomedical images at multi-scale resolutions using statistical modeling, computer vision, machine learning, and graph-based deep learning techniques. He has developed explainable artificial intelligence (AI) algorithms to understand the origins of diagnostic discordance in differentially diagnosing a broad spectrum of breast lesions from digitized histopathology images. He is also serving as a Member of the Review Board for Journal of Pathology Informatics (JPI) and Signal, Image and Video Processing. He is also an active member of the Digital Pathology Association (DPA).
Link to the papers being discussed :
1. pitt.app.box.com/s/nk7ll6r1yt... (Paper accepted at MICCAI-2020)
2. link.springer.com/chapter/10.... (Paper accepted at MICCAI-2021)
Reach out to the Speaker :
akashparvatikar.github.io
/ akash007