Pancreas segmentation in CT scans: A novel MOMUNet based workflow.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Automatic pancreas segmentation in CT scans is vital for medical applications but faces challenges.
• Pancreas segmentation is difficult due to size variations and class imbalance in CT scans.
• Existing methods struggle with accuracy and computational efficiency.

Purpose of the Study:

• To develop a novel, efficient, and accurate three-stage workflow for automatic pancreas segmentation.
• To address challenges of class imbalance and anatomical variations in CT scans.
• To create an ultra-lightweight deep learning model for low-resource medical facilities.

Main Methods:

• External Contour Cropping (ECC) to mitigate class imbalance.
• Size Ratio (SR) technique to improve model robustness against anatomical variations.
• MOMUNet, an ultra-lightweight segmentation model (1.31M parameters).

Main Results:

• Achieved a 2.56% Dice Score (DSC) improvement on NIH-Pancreas dataset.
• Achieved a 2.97% DSC improvement on MSD-Pancreas dataset.
• Demonstrated 68.4% DSC for colon cancer segmentation, surpassing SOTA.

Conclusions:

• The proposed workflow significantly improves segmentation accuracy for small abdominal organs like the pancreas and colon.
• The approach enhances computational efficiency, making advanced deep learning accessible for low-resource settings.
• This work advances automated medical image analysis for improved diagnosis and treatment planning.