The ASC Module: A GPU Memory-Efficient, Physiology-Aware Approach for Improving Segmentation Accuracy on Poorly...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Aging populations, like Japan's, risk inadequate medical resources.
• Accurate pre-surgical aorta localization via CT is crucial.
• Current AI models struggle with poorly contrasted CT images.

Purpose of the Study:

• To enhance AI-driven aorta segmentation on low-contrast CT scans.
• To improve pre-surgical planning accuracy.
• To address limitations of existing semantic segmentation models.

Main Methods:

• Developed an Automatic Spatial Contrast (ASC) Module.
• Integrated ASC Module with UNet, Attention UNet, TransUNet, and Swin-UNet.
• Evaluated model performance on aorta segmentation in CT images.

Main Results:

• Significant improvements in Intersection-over-Union (IoU) up to 24.84%.
• Substantial gains in Dice Similarity Coefficient (DSC) up to 28.13%.
• Minimal increase in GPU memory usage compared to baseline models.

Conclusions:

• The ASC Module effectively improves aorta segmentation accuracy on challenging CT images.
• This AI enhancement aids pre-surgical planning, particularly for at-risk populations.
• The method offers a practical solution with efficient resource utilization.