Title: 

Regression-based approach for predicting genetic mutations from tumour pathology slides 

Abstract: 

This research explores the potential of deep learning techniques to predict genetic mutations from Hematoxylin and Eosin (H&E) stained Whole Slide Images (WSIs) of tumor tissues. Utilizing the CPTAC-BRCA (Clinical Proteomic Tumor Analysis Consortium Breast Invasive Carcinoma) dataset, which includes WSIs and gene expression data from 106 breast cancer patients, two models were constructed:(i) a direct model that performs binary classification (mutated vs. non-mutated) for each gene based on the extracted features from the images, and (ii) an indirect model that uses a regression approach to predict the expression levels of 18,272 genes and then determines gene mutation status based on the median expression levels of these genes in the whole population, calculated from the TCGA dataset. The indirect model showed superior accuracy in genetic mutation prediction and gene expression prediction. We also implemented attribution analysis using the Integrated Gradients method to generate heatmaps, highlighting regions within the WSIs that significantly influence the prediction outcomes. These heatmaps assist in identifying critical areas affecting tumor pathology, offering insights for enhanced diagnostic accuracy and treatment strategies. 

Biography: 

Ke Li, a master student of mathematical sciences at the Mathematical Sciences Institute (MSI) of the Australian National University (ANU), studies in computational pathology leveraging deep learning methods. Ke investigates how technological advancements can enhance cancer diagnosis and treatment, assisting medical professionals in decision-making and providing effective references.