A Novel Group Block Attention Module for Lung CT Image Segmentation

Yuchai Wan, Pengfei Ren, Yifan Li, Lili Zhang, Murong Wang, Ruijun Liu

Abstract: Computed Tomography (CT) imaging holds a crucial role in diagnosing lung diseases. Employing computer vision to segment lesions in CT scans aids physicians in precise, effective diagnoses. However, it is challenging for traditional convolution neural network(CNN) based methods to segment small lung lesions accurately. To address this issue, in this paper, we propose a novel attention mechanism, namely the Group Block module. The Group Block uses unsupervised grouping idea, and allocates the image block tokens in the up-down sampling module to different semantic category tokens. Image block tokens are allocated to each group token and merged into a larger group with higher-level semantic information. The grouping token is handled in a similar way to the self-attention mechanism. The comparison experiments, with and without the Group Block module, validated the effectiveness and strong generalization ability of our method.