Weakly Supervised Learning for Medical Image Segmentation

Abstract

Medical image segmentation plays a critical role in clinical diagnosis and treatment planning; however, the performance of deep learning models typically depends on large quantities of pixel-wise annotated data, which are costly and time-consuming to obtain. This thesis investigates annotation-efficient approaches for medical image segmentation by developing weakly supervised learning frameworks that reduce annotation requirements while maintaining high segmentation accuracy.

The first contribution proposes an enhanced scribble-based segmentation framework that extends a modified U-Net architecture with increased feature-learning capacity, mix-based data augmentation, and global and local consistency regularisation. These components improve learning from sparse scribble annotations and promote anatomically coherent segmentation masks. Experimental results demonstrate improved segmentation performance compared with existing scribble-supervised methods.

The second contribution presents an ensemble learning framework for weakly supervised medical image segmentation based on single bounding-box annotations. The proposed framework integrates U-Net, ResUNet, and DeepLabV3 using bagging and boosting strategies to exploit complementary architectural characteristics. Despite relying solely on coarse single-bounding-box supervision, the ensemble models improve robustness, reduce model-specific bias, and yield more stable segmentation predictions. Experimental results demonstrate that the ensemble methods consistently outperform the individual models, with boosting achieving the highest segmentation performance.

The third contribution introduces a novel dual bounding-box supervision strategy that utilises an inner bounding box to represent high-confidence object regions and an outer bounding box to provide contextual anatomical information. A region-aware weighted loss function encourages accurate object localisation while reducing ambiguity in boundary regions. The proposed framework significantly outperforms conventional single-box weakly supervised approaches and narrows the gap between weak and full supervision.

Overall, the findings demonstrate that high-quality medical image segmentation can be achieved using sparse scribble annotations, ensemble learning techniques, and coarse bounding-box supervision without requiring exhaustive pixel-wise labels. The proposed methods reduce annotation effort while maintaining competitive segmentation accuracy, providing a practical and scalable pathway for deploying deep learning-based segmentation systems in clinical environments.