Dual-Attention Model Fusing CNN and Transformer for Pancreas Segmentation.

Abstract

Pancreatic cancer is a highly malignant cancer of the digestive tract and is rapidly progressing and spreading clinically. Automatic and accurate pancreatic tissue segmentation in abdominal CT images is essential for the early diagnosis of pancreatic-related diseases. It is challenging that the pancreas is small in size and complex in morphology. To address this problem, we propose a dual-attention model fusing CNN and Transformer to effectively activate pancreas-related features expression. The CNN structure weights the importance of pancreas-related features at the channel level and weakens the background information. Transformer feature aggregation module constructs spatial correlations among long-distance pixels from a global perspective. This study is validated on the NIH-TCIA dataset and achieved a mean Dice Similarity Coefficient of 85.82%, which is outperforming than the state-of-the-art methods. The visualization of surface distance also demonstrates the effective segmentation of pancreas boundary details by the proposed model.