CAFU-Net: A Context-Aware Feature Aggregation Network for Lung Nodule Segmentation

Accurate segmentation of pulmonary nodules from CT images is of significant importance for the early diagnosis and treatment of lung cancer and related diseases. However, the segmentation task faces challenges due to the overlap between pulmonary nodules and surrounding anatomical structures, as well as variations in size, shape, and density. These challenges are further compounded by low contrast and complex structural backgrounds. To address these issues, this paper proposes an improved U-Net network—CAFU-Net. The network significantly enhances the accuracy and robustness of pulmonary nodule segmentation by integrating context-aware information based on a gating mechanism, a shared-weight triple attention mechanism, and a dual-scale selective convolution kernel structure. Quantitative and qualitative experiments were conducted on the publicly available LIDC (The Lung Image Database Consortium) dataset and the private MID-FAHGMU (The Fourth Affiliated Hospital of Guangxi Medical University Medical Imaging Department) dataset. The experimental results show that, compared to 10 advanced medical image segmentation methods, CAFU-Net achieves an IoU of 86.51% on the LIDC dataset, surpassing the best-performing comparative method, BRAU-Net++, by 0.55%. The Dice coefficient is 91.71%, the F1-score is 91.78%, the F2-score is 91.78%, and the F0.5-score is 92.39%, all significantly outperforming other methods. On the MID-FAHGMU dataset, CAFU-Net achieves an IoU of 63.52%, a Dice coefficient of 76.21%, an F1-score of 76.47%, an F2-score of 77.06%, and an F0.5-score of 77.74%, exceeding most comparative methods in several metrics. These experimental results fully validate the superiority and robustness of CAFU-Net in the task of pulmonary nodule segmentation. The source code and segmentation images will be released at https://github.com/XiaoBai-Lyq/CAFU-Net.