Relieve the dilemma of generalization in machine learning for medical images

Abstract

Medical imaging often relies on small, private datasets due to limited data availability, high annotation costs, and patient privacy constraints. Furthermore, patient variability, inter-dataset differences, and pathological conditions introduce substantial heterogeneity, hindering the generalization and reproducibility of machine learning (ML) models. This thesis focuses on two specific tasks—skull stripping and tumor subregion identification—to investigate how incorporating geometric or clinical priors into ML model design can alleviate the impact of heterogeneity and improve robustness, generalizability, and real-world applicability.

First, for skull stripping, we draw inspiration from traditional template-based methods that utilize average brain templates. We introduce a Geometrically Enhanced Convolutional Neural Network (G-CNN) to effectively extract and leverage brain template priors, mitigating the adverse effects of input heterogeneity. We further advance this concept by developing the Geometric-Information-Assisted Network (GINet) to provide patient-specific geometric information (GI) to better learn and utilize geometric priors of brain structure.

Second, for tumor subregion identification, clinicians frequently rely on unsupervised clustering methods to define distinct tumor habitats due to the absence of a standardized criterion. However, patient heterogeneity reduces the reproducibility and robustness of clustering-based models. To address this challenge, we propose a clinical-prior-based autoencoder integrated with a Bayesian optimization (BO) framework, along with a novel clustering stability evaluation metric. We further extend this work to a multi-objective Bayesian optimization (MOBO) framework that balances clustering stability with clinical relevance. Rigorous statistical analyses validate the efficacy of our proposed frameworks, demonstrating enhanced reproducibility and stable tumor subregion identification across heterogeneous datasets.

By incorporating geometric or clinical priors into ML models, this thesis establishes a pathway toward developing robust, reproducible, and reliable models applicable to diverse medical imaging scenarios.

Date of Award	2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Xi Chen (Supervisor) & Tom Fincham Haines (Supervisor)