Subset selection strategy-based pancreas segmentation in CT.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Convolutional Neural Network (CNN)-based methods are prevalent in medical image segmentation.
• Imbalance problems in medical image data reduce segmentation accuracy and validity.
• Existing CNN methods struggle with foreground-to-background and hard-to-easy data imbalances.

Purpose of the Study:

• To address data imbalance issues in medical image segmentation.
• To enhance the accuracy and reliability of CNN-based segmentation models.
• To improve the segmentation of challenging medical images.

Main Methods:

• Proposed positive-negative subset selection (PNSS) to balance foreground-to-background data.
• Introduced hard-easy subset selection (HESS) to manage difficult-to-segment samples.
• Implemented methods to enhance model focus on challenging image features.

Main Results:

• Significantly improved segmentation accuracy in the worst-case scenarios.
• Achieved nearly 5% increase in minimum Dice Similarity Coefficient (DSC) on pancreatic segmentation.
• Demonstrated robust performance gains across liver, liver tumor, and brain tumor segmentation tasks.

Conclusions:

• PNSS and HESS effectively alleviate data imbalance problems in medical segmentation.
• The proposed methods enhance segmentation accuracy, especially for difficult samples.
• The techniques ensure the reliability of medical image segmentation for clinical applications.