X-LAT-Net: An Interpretable Lightweight Axial Transformer Network for Pancreatic CT Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Pancreas segmentation in CT images is challenging due to anatomical depth, morphological variations, and low contrast.
• Deep learning models improve segmentation but often lack efficiency and interpretability, hindering clinical trust and deployment.

Purpose of the Study:

• To develop X-LAT-Net, a lightweight network for efficient, accurate, and interpretable pancreas segmentation in CT images.
• To address the limitations of current deep learning models in terms of computational complexity and lack of transparency.

Main Methods:

• Designed a U-shaped architecture incorporating an Axial Depthwise Convolution module for efficient long-range dependency capture.
• Introduced an Interpretable Cross-scale Transformer (X-CATrans) for global context modeling and generating attention heatmaps for interpretability.
• Utilized a Shift-enhanced MLP module to refine the segmentation of indistinct pancreatic boundaries.

Main Results:

• X-LAT-Net achieved a Dice coefficient of 82.34% on the NIH Pancreas-CT dataset with only 1.6 million parameters.
• Outperformed mainstream methods in accuracy and inference speed.
• Demonstrated visual interpretability through attention heatmaps, enhancing clinical confidence.

Conclusions:

• X-LAT-Net offers an efficient, accurate, and interpretable solution for pancreas segmentation.
• The model's lightweight design and interpretability make it suitable for resource-constrained clinical settings and AI-aided pancreatic cancer screening.