M³Net: A multi-scale multi-view framework for multi-phase pancreas segmentation based on cross-phase non-local...

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Distinguishing the pancreas from surrounding structures in CT scans is challenging.
• Utilizing complementary information from arterial and venous phases of CT scans can improve segmentation.
• Current computer-aided pancreas segmentation methods have limited exploration of cross-phase contextual information.

Purpose of the Study:

• To present M³Net, a framework for multi-phase pancreas segmentation that integrates multi-scale, multi-view, and cross-phase information.
• To improve the accuracy and robustness of computer-aided pancreas segmentation by leveraging complementary visual data from different CT phases.
• To develop attention modules that enhance feature representation and handle potential misalignments across phases.

Main Methods:

• Developed M³Net, a dual-path network with individual branches for two CT phases (arterial and venous).
• Incorporated cross-phase interactive connections to integrate complementary visual information between the two phases.
• Designed location attention and depth-wise attention modules to enhance high-level feature representation and suppress misalignment.

Main Results:

• Achieved state-of-the-art performance with an average Dice Similarity Coefficient (DSC) of 91.19% on internal CT data.
• Demonstrated promising results with an average DSC of 86.34% on external CT data.
• The proposed attention modules effectively improved feature representation and handled cross-phase correlations.

Conclusions:

• M³Net effectively integrates multi-scale, multi-view, and cross-phase information for accurate pancreas segmentation.
• The framework shows significant potential for improving computer-aided diagnosis in medical imaging.
• The developed attention mechanisms contribute to robust feature extraction and segmentation performance across different datasets.