Automated segmentation of lymph node metastases in colorectal cancer based on convolutional KAN

Accurate preoperative segmentation of colorectal cancer lymph node metastasis holds significant clinical value for formulating individualized surgical strategies. However, accurate segmentation remains a formidable challenge due to the complex backgrounds, small sizes, irregular shapes, and indistinct boundaries typically exhibited by colorectal cancer lymph nodes. Consequently, this paper proposes the U-Convolutional KANs Attention Net (UCKAN). First, in complex environment, we replace the convolutional layers with the convolutional KAN module, leveraging the learnable non-linear fitting advantage of KAN to enhance the ability to accurately extract the features of small targets. Second, we invented the Global Spatial Multi-scale Attention (GSMA) module, which integrates global context and multi-scale features to improve the model's sensitivity to small-volume lymph nodes, achieving precise spatial focusing on lesions. Evaluations on practical medical imaging tasks show that the UCKAN model has superior performance., with Dice and IoU scores reaching 91.52 and 83.29 %, respectively, comprehensively outperforming eight advanced medical image segmentation models. The innovative mechanisms proposed in this study effectively address the segmentation challenges associated with low-contrast and multi-scale lesions, providing a reliable quantitative basis for formulating individualized preoperative surgical strategies and demonstrating significant potential for clinical translation.