Semi-supervised breast ultrasound image segmentation via multi-paradigm synergy and cross-cyclic distillation

Medical image segmentation methods based on deep learning often require a large amount of labeled data to achieve better segmentation performance. However, acquiring a wealth of labeled images in medical scenarios is still a challenging. Semi-supervised learning (SSL) alleviates this dependency by utilizing unlabeled samples, improving data utilization. In this paper, we propose a novel SSL method based on Multi-paradigm Synergy and Cross-cyclic Distillation (MSCD) to effectively leverage unlabeled images. Firstly, we introduced a Transformer structure in the deeper layers of the encoder, combining CNN and Transformer paradigms to encourage the network to enhance feature learning through multi-paradigm synergy. Secondly, we proposed a Multi-paradigm Feature Interaction Module (MFIM) to achieve the fusion and interaction of CNN and Transformer features, and to promote the learning of global features in the CNN network. Subsequently, we employed a three-decoder structure to encourage consistency between decoders by cross-cyclic knowledge distillation and multi-layer cross consistency strategy. Finally, extensive experiments with several state-of-the-art SSL segmentation methods for medical images were performed on publicly available breast ultrasound datasets. The experimental results showed the effectiveness of our method and its better performance in breast lesion segmentation.