Weakly Supervised Learning of Deep Neural Networks for Histology

General presentation of the topic:
Cancer remains a leading cause of mortality in Canada and worldwide. Although histopathological classification of cancer has been a mainstay in the clinical setting, the emergence of omics data has led to a shift towards molecular classification, which is associated with varying responses to treatment and survival outcomes. However, much work remains to identify and characterize tumor heterogeneity accurately. We propose integrating multi-OMICS within the context of histological data to provide a more informative view of tumor characteristics, which results in better predictive models of patient outcomes. However, techniques that integrate histopathological and molecular classification still need to be developed.
Moreover, state-of-the-art deep learning (DL) models for cancer assessment present several challenges related to the heterogeneous nature and availability of annotated data. This project builds on our successful collaboration bridging AI with oncology. By leveraging the multi-modal fusion of multi-omics and histology data, we will develop robust, deep, weakly-supervised learning (WSL) models that do not require costly pixel-wise annotation by experts. These models will yield accurate context-aware predictions and high-resolution attention maps that favor the interpretation of salient cancer regions for survival outcomes. Our proposed project will leverage genomic (mutation, CNA) and transcriptomic datasets from The Cancer Genome Atlas and corresponding histological images from the Cancer Digital Slide Archive. Epigenetic and proteomic data will be incorporated into the models as available. This project will provide new methods to the research community and open-access software tools for diverse end-users.

Objectives:
This project will investigate and develop DL models that integrate Omics data and corresponding histology image data to address three primary objectives: A) Develop fusion models incorporating multi-Omics datasets with image analytics to predict patient outcomes; B) Establish pixel-wise localization that projects genomic heterogeneity onto histological images; C) Obtain visual interpretability feedback for the model’s decisions. Current DL models for cancer assessment present several challenges related to the heterogeneous nature and availability of annotated data.