Contrastive Registration for Unsupervised Medical Image Segmentation
This work provides a strong advancement in unsupervised medical image segmentation, addressing the critical need for automated and annotation-free methods for clinicians and researchers in the medical domain, which is an incremental improvement over existing unsupervised techniques.
This paper tackles the problem of unsupervised medical image segmentation, which is crucial for diagnosis but hindered by the high cost and bias of manual annotations. The authors propose a novel CNN-based contrastive registration architecture that exploits both image-level registration and feature-level contrastive learning to achieve registration-based segmentation. Their method substantially outperforms current state-of-the-art unsupervised segmentation methods on two major medical image datasets.
Medical image segmentation is a relevant task as it serves as the first step for several diagnosis processes, thus it is indispensable in clinical usage. Whilst major success has been reported using supervised techniques, they assume a large and well-representative labelled set. This is a strong assumption in the medical domain where annotations are expensive, time-consuming, and inherent to human bias. To address this problem, unsupervised techniques have been proposed in the literature yet it is still an open problem due to the difficulty of learning any transformation pattern. In this work, we present a novel optimisation model framed into a new CNN-based contrastive registration architecture for unsupervised medical image segmentation. The core of our approach is to exploit image-level registration and feature-level from a contrastive learning mechanism, to perform registration-based segmentation. Firstly, we propose an architecture to capture the image-to-image transformation pattern via registration for unsupervised medical image segmentation. Secondly, we embed a contrastive learning mechanism into the registration architecture to enhance the discriminating capacity of the network in the feature-level. We show that our proposed technique mitigates the major drawbacks of existing unsupervised techniques. We demonstrate, through numerical and visual experiments, that our technique substantially outperforms the current state-of-the-art unsupervised segmentation methods on two major medical image datasets.