Medical Image Segmentation under Data-Efficient Scenarios

Abstract

Medical image analysis is essential for accurate diagnosis, treatment planning, and patient management. Among its tasks, segmentation is key to identifying anatomical structures and abnormalities. Although deep learning has significantly advanced segmentation by learning complex patterns from large datasets, it requires vast amounts of labeled data, which is expensive and time-consuming to obtain. Consequently, transfer learning has emerged to bridge distribution gaps and reduce reliance on extensive labeled data. Nonetheless, challenges remain, such as semantic gaps in image categories, limitations of adversarial learning, and issues with generalizing to unseen target data. This thesis presents several deep learning-based domain adaptation methods for data-efficient medical image segmentation:

The first part of this thesis introduces CAPL-Net, a novel deep unsupervised domain adaptation (UDA) approach specifically designed for cross-domain nuclei instance segmentation and classification using category-aware feature alignment and pseudo-labeling.

The second contribution presents an advanced dynamic diffusion-based framework (DDF-UDA) for cross-domain optic disc (OD) and optic cup (OC) segmentation in fundus images. DDF-UDA includes an adaptive module based on feature- and pixel-level diffusion processes, along with a Nash equilibrium-based adjustment strategy to reduce cross-domain discrepancies.

Finally, we propose a novel zero-shot domain adaptation (ZSDA) paradigm, named LMaD, to address the challenge of requiring access to target images during training. LMaD utilizes a language model as a feature disentangler for ZSDA medical image segmentation tasks, aiming to eliminate interference caused by domain-specific information, resulting in superior performance.

Overall, this thesis advances UDA and ZSDA methods in medical image segmentation, offering practical solutions to overcome data scarcity and enhance model generalization in diverse clinical settings.