Rethinking U-Net architecture in medical imaging: Advancing the efficient and interpretable UKAN-CBAM framework for...

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computational Pathology
• Oncology

Background:

• Prompt detection of colorectal polyps is crucial for preventing colorectal cancer.
• Manual polyp detection is challenged by high costs, expertise requirements, and error susceptibility.
• Existing methods often lack interpretability and efficiency.

Purpose of the Study:

• To develop an advanced semantic segmentation framework, UKAN-CBAM, for enhanced colorectal polyp detection.
• To integrate Kolmogorov-Arnold Networks (KANs) with Convolutional Block Attention Modules (CBAM) for improved performance.
• To achieve a computationally efficient and interpretable model for clinical applications.

Main Methods:

• Proposed UKAN-CBAM framework combining KANs and CBAM within a U-Net architecture.
• Training on the Kvasir-SEG dataset and validation across multiple external datasets.
• Utilized 10-fold cross-validation and Grad-CAM for robustness and interpretability analysis.

Main Results:

• UKAN-CBAM achieved superior performance, outperforming SOTA methods with mDice of 93.80% and mIoU of 89.18%.
• Demonstrated computational efficiency with 55.99 MB memory usage and 122.272 ms inference speed.
• Statistical significance confirmed through paired t-tests and cross-validation, highlighting model robustness.

Conclusions:

• UKAN-CBAM offers an effective and reliable tool for real-time clinical applications in colorectal polyp detection.
• The integration of attention mechanisms and interpretability marks a significant advancement in medical diagnostics.
• The framework shows strong generalization capabilities across diverse datasets.