Dual-Attention Model Fusing CNN and Transformer for Pancreas Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Pancreatic cancer is a dangerous digestive tract malignancy with rapid progression.
• Accurate segmentation of pancreatic tissue in abdominal CT images is vital for early diagnosis.
• The pancreas's small size and complex morphology present segmentation challenges.

Purpose of the Study:

• To develop an advanced model for automatic and accurate pancreatic tissue segmentation in CT images.
• To enhance the expression of pancreas-related features by fusing Convolutional Neural Network (CNN) and Transformer architectures.
• To overcome the limitations of segmenting small and morphologically complex pancreatic structures.

Main Methods:

• Proposed a dual-attention model integrating CNN and Transformer.
• CNN component focuses on channel-level feature weighting for pancreas identification and background suppression.
• Transformer component aggregates spatial correlations among pixels globally.

Main Results:

• Achieved a mean Dice Similarity Coefficient (DSC) of 85.82% on the NIH-TCIA dataset.
• Demonstrated superior performance compared to state-of-the-art segmentation methods.
• Surface distance visualizations confirmed effective segmentation of pancreas boundary details.

Conclusions:

• The proposed dual-attention model effectively segments pancreatic tissue in abdominal CT images.
• The fusion of CNN and Transformer architectures enhances feature expression and spatial correlation.
• This method shows significant potential for improving early diagnosis of pancreatic diseases.