Semi-supervised ultrasound image segmentation via dual-backbone network and bidirectional contrastive learning

Semi-supervised semantic segmentation is crucial for medical ultrasound image analysis, but it faces challenges due to limited labeled data and difficulty in capturing global and local features, and traditional contrastive learning often struggles with pixel-level tasks because it cannot capture fine spatial relationships within an image. To address this issue, we propose a dual-backbone network that combines convolutional neural networks (CNNs) for global feature extraction and Transformers for local context understanding. Our method adopts a bidirectional contrastive learning strategy to enforce feature consistency and integrates channel and spatial attention mechanisms to focus on key image regions. We evaluate our method on breast ultrasound image (BUSI) and ovarian ultrasound datasets, achieving higher segmentation accuracy than existing methods.