Deep Learning for Medical Image Segmentation

Medical image segmentation (MIS) plays a fundamental role in a wide range of medical image analysis applications. Over the past decade, MIS has been significantly advanced by deep learning (DL). However, the effectiveness of DL-based methods typically depends on large-scale, precisely annotated datasets, which are often unavailable in the medical domain due to the high cost and expertise required for medical data acquisition and annotation. When trained on limited data, deep neural networks (DNNs) are prone to over-fitting, resulting in poor generalisation performance when deployed on previously unseen data. In MIS, this generalisation gap is exacerbated by two key challenges: 1) data scarcity: due to the limited quantity of annotated images available for training, over-fitting the training data causes poor generalisation performance on images unseen during training; 2) expert-specific annotation: variability in annotations across experts arises from the inherent complexity of medical images and differences in expert knowledge. Consequently, a model trained on certain experts' annotations may under-perform when deployed in clinical contexts with new images and new experts. This thesis aims to address these domain-specific limitations, with an overarching goal of generating enhanced supervision signals to improve the generalisation capability of DNNs for MIS across both images and user experts under constrained data settings. Toward achieving this goal, our contributions are threefold.First, we tackle the expert-specific annotation problem by leveraging the supervision from multiple experts via multi-expert annotation, aiming to achieve adaptive MIS that generalises to new experts' annotations. Through extensive empirical analysis, we investigate when and how multi-expert annotation improves a DNN model's generalisation capability to new experts' annotations and what DNN training method can best achieve such generalisation. Our findings provide detailed evidence for the efficacy of multi-expert annotation and practical insights into developing generalisable models that can efficiently adapt to new experts in the presence of expert-specific annotations.Second, we intend to mitigate the impact of data scarcity and improve the model's generalisation to new images by leveraging self-supervision.We develop an adaptive training method, which generates diverse self-supervised tasks from the existing annotated dataset, dynamically selects the most beneficial combination of tasks, and adaptively transfers the knowledge learnt from them to enrich the segmentation model's knowledge.Compared to existing approaches, our method integrates more diverse self-supervised tasks and offers superior effectiveness and efficiency in exploiting their supervision signals to benefit segmentation performance.The third study moves beyond the single-site setting and generates enhanced supervision via multi-site collaboration, where the goal is to improve generalisation performance for participating sites while navigating strict privacy constraints and potential security threats arising from non-transparent local processes. We propose a new federated learning method, tailored for the multi-site MIS scenario, to enable privacy-preserving defence against poisoning attacks by sustaining robust performance and localising the compromised site. Compared to previous approaches, our method provides utility-based defence and diagnosis, does not depend on public datasets, and works for the setting with few participating sites. These features enable more practical and effective federated learning with enhanced privacy preservation and security for multi-site MIS.Extensive experiments across diverse datasets and clinical settings demonstrate the effectiveness of the proposed methods in improving DNN model generalisation for MIS tasks. Collectively, this thesis contributes practical methodologies to address domain-specific challenges in MIS and advances the practical deployment of DNNs in real-world clinical environments.