Pancreas Segmentation Based on Multi‐Stage Attention Enhanced U‐Net

ABSTRACT Considering that the pancreas occupies a very small proportion of the abdominal organs and varies in size, shape, and position, U‐Net encounters challenges related to intra‐class inconsistency and inter‐class indistinction in pancreatic segmentation tasks. To address these issues, this paper proposes a pancreas segmentation method based on a multi‐stage attention enhanced U‐Net to effectively leverage feature information at each stage of the U‐Net architecture. In particular, during the encoding phase of U‐Net, triple attention is utilized to capture dependencies between different dimensions; during the skip connection phase, a channel cross‐fusion Transformer is introduced to fuse multi‐scale channel information from different layers of the encoder; and during the decoding phase, feature integration convolution is employed to enhance the model's capacity for integrating global and local information. A four‐fold cross‐validation was performed on 82 Three‐Dimensional Computed Tomography (3D CT) scans from the National Institutes of Health (NIH) and 281 3D CT scans from the Medical Segmentation Decathlon (MSD) to evaluate the proposed model. Experimental results demonstrate that the proposed method achieved superior performance on both pancreatic datasets, surpassing mainstream pancreatic segmentation methods, with average Dice scores of 87.16% and 87.53%, respectively, yielding improvements of 2.58% and 1.71% compared to U‐Net. The proposed method is an end‐to‐end pancreatic segmentation algorithm suitable for small organ region segmentation in complex tissues, capable of high‐precision pancreatic segmentation in processing entire 3D CT image slices.