Deep Neural Networks Integrating Functional Genomics and Histopathological Images Data for Predicting Stages in Colon Cancer

Abstract

A deep neural network (DNN) is a type of artificial neural network (ANN) that uses multiple layers of interconnected nodes to process data and solve complex problems. Successful deep neural network learning model can identify complex or non-linear relationships in data and extracts most informative feature for improved prediction accuracy. This technique is desirable in colon cancer staging and survival risk stratification using single or multi-omics data, wherein even the state-of the- art model find it challenging to discern informative knowledge from the plethora of features in diverse biological data. This challenges is due to the heterogeneity and complexity of cancer disease of the same stages in different samples, unbalance sample size, limited data and several other factors. Deep neural network learning model is an evolving and emerging field with potential of alleviating this challenges but output from this model are seen as a Blackbox due to unexplainable or lack of interpretation nature of its decision making techniques.
This thesis aims to develop a robust deep neural network learning framework that incorporates contextual explainability for effective knowledge extraction and information analysis from multiomics datasets with complex non-linear features. The proposed interpretable models were implemented and evaluated on omics datasets for tasks such as cancer stages prediction and survival risk stratification. Furthermore, the performance of these models was compared with related studies to demonstrate their effectiveness in achieving a reliable and interpretable predictive model for colon cancer. Our first contribution is the aggregation of atypia patterns in Haematoxylin and Eosin (H&E) histopathological images with diverse carcinogenic expression from messenger Ribonucleic acid (mRNA), micro Ribonucleic acid (miRNA) and Deoxyribonucleic (DNA) methylation as an integrative input source into a deep network for colon cancer stages classification and samples stratification into low or high-risk survival groups. Our second contribution is the design and implementation of extended Squeeze-and-Excitation Multiheaded Attention (ESEMA) model which efficiently integrates and enhances the multimodal features, capturing biological pertinent patterns that improves both the accuracy and explainability of cancer stages predictions. Also, a portion of the developed framework called AMAStage (Attention and Multi-Autoencoder: Explainable multi-omics feature selection and cancer staging) perform favorably well compared with other related state-of-the-art(SOTA) model.

Date of Award	19 Feb 2026
Original language	English
Awarding Institution	Northumbria University
Supervisor	Wai Lok Woo (Supervisor) & Zeyneb Kurt (Supervisor)