PC-VMamba: Position and channel-guided vision mamba for small object segmentation in medical images

Automatic segmentation of small objects in medical images can efficiently and objectively identify tiny but critical lesions in medical images, enabling ultra-early diagnosis of diseases and providing valuable opportunities for treatment. However, existing methods still suffer from lower segmentation accuracy and poor generalization ability. To address this issue, we propose a new Position and Channel-guided Vision Mamba framework (PC-VMamba) to enhance segmentation accuracy and generalization capabilities for small target lesions. Building on the long-range modeling capabilities and linear computational complexity advantages of Vision Mamba (ViM), the proposed PC-VMamba introduces a dual attention module, which combines the position feature and channel extraction capabilities of the position attention module (PAM) and the channel attention module (CAM), respectively. Meanwhile, it adopts a learnable cross-scan strategy to achieve more comprehensive feature modeling capabilities, thereby improving segmentation accuracy and stability. Besides, multiple skip connections are established between the encoder and decoder to bridge the semantic gap between high-level semantic features and low-level detail features, thus improving the detection capability of lesion boundaries and small targets. We tested the proposed method on three public datasets(BUSI, ISIC2017, LUNA16). Experimental results show that the proposed PC-VMamba significantly improves segmentation accuracy compared to state-of-the-art methods while effectively controlling parameter complexity. Furthermore, generalization experiments on the Kvasir, CVC-ClinicDB and REFUGE2 public datasets further demonstrate the robustness of the proposed method in medical image segmentation of small objects and blurry boundaries. Our code can be found at https://github.com/Weixuyang27/PC-VMamba .