Fuzzy-based cross-image pixel contrastive learning for compact medical image segmentation

Existing medical image segmentation ignore the exploration of inter-class similarity and intra-class variability in pixel semantics, and aim to develop deeper and more complex networks for strength enhancement, leading to insufficient pixel relationship modeling high computational cost. To overcome the aforementioned limitation, we propose a novel fuzzy-based cross-image pixel contrastive learning regime to exploit discriminative relationships between pixel representations across images globally. CPC ensures that the lesion pixel is pulled closer to other lesion pixels while pushed far away from the background pixels in the representation space, thus driving the network to discriminate pixel semantics more robustly. Instead of computing or storing all samples, we devise a fuzzy filtering strategy that selects Top-K samples based on fuzzy membership. Furthermore, considering the speed requirement of medical image segmentation, we propose a compact but efficient network for rapid and precise segmentation, which can model both local and long-range dependencies by microscopically fusing Transformer and convolution. Benefitted from our efficient design of the hybrid module, the proposed network enjoys the properties of being compact, lightweight, and powerful. We term our efficient hybrid network with cross-image pixel contrastive learning as CPCNet. Extensive qualitative and quantitative experiments on various image segmentation tasks demonstrate that our CPCNet surpasses the state-of-the-art approaches.