Deep Learning-based Tumour Heterogeneity Analysis with Multiparametric Magnetic Resonance Imaging

Abstract

Tumour heterogeneity, characterised by spatial and temporal variations in cellular morphology and molecular profiles within the tumour tissues, impacts oncology diagnosis, prognosis, and treatment outcomes. Medical image analysis, particularly through multiparametric magnetic resonance imaging (mpMRI), provides a non-invasive approach to capture these complex heterogeneity patterns, enabling a more objective and comprehensive assessment of tumour biology. However, the intrinsic variability in heterogenous tumours poses challenges in accurately delineating tumour regions and predicting therapeutic responses, often leading to inconsistent clinical interpretations. Because of this, labelling and analysing heterogeneity sub-regions in mpMRI are time-consuming tasks and requires experienced expertise.

This thesis introduces a novel deep-learning framework that addresses the challenges of tumour heterogeneity in mpMRI modalities to enhance tumour heterogeneity analysis in medical image and its use in downstream tasks including image segmentation and classification. The framework comprises of two main components. First component is an unsupervised semantic segmentation method developed to delineate tumour sub-regions automatically. This method effectively captures the intrinsic structure of heterogeneous tumours by leveraging a multi-phase training strategy that combines coarse segmentation with refined, self-supervised learning enhanced by sparse spatial continuity and context-based hierarchical loss functions. Second, we propose a heterogeneity-aware deep learning method for tumour classification that integrates machine-generated sub-region labels with dual-stream feature extraction for both local heterogeneity and global image information. A learnable alignment module is employed to standardise sub-region labels across different imaging modalities, enabling the extraction of both local heterogeneity features and global contextual information.