CDSegNet: a multi-scale convolution and attention mechanism U-shaped network for Crohn's disease lesion segmentation

Accurate segmentation of Crohn's disease (CD) lesions from computed tomography enterography (CTE) cross-sectional images is crucial for diagnosing CD patients and may assist in developing a personalized treatment plan. However, to the best of our knowledge, studies on automatic CD lesion segmentation remain limited. This paper proposes a novel model (named CDSegNet) based on the U-Net to improve the segmentation of CD lesions. The model integrates a residual dilated and standard convolution feature extract (RDCFE) module to enhance fine-grained and global feature extraction while preserving information flow. Furthermore, a residual attention feature extraction (RAFE) module is introduced in the decoder to refine features and sharpen ambiguous lesion boundaries. In addition, a multi-scale convolution module is designed to aggregate features from different receptive fields to improve robustness across varying lesion sizes. Finally, Intersection over Union (IoU), Recall, Dice Similarity Coefficient (DSC), and Hausdorff distance (HD) are employed to quantitatively assess the model performance. The Recall, IoU, HD, and DSC of the CDSegNet are 0.867, 0.820, 23.28, and 0.895, respectively. Compared to the baseline model U-Net on our dataset, the IoU, HD, and DSC improve by 11.6%, 3.21, and 10.4%, respectively. Experiment results demonstrate that CDSegNet exhibits competitive performance under the current experimental setting, with particular strengths in segmenting small and medium lesions and achieving stable segmentation across all test images. However, further validation is needed prior to clinical application.