Uncertainty-Guided Adaptive Correction for Semi-Supervised Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semi-supervised learning is key for medical image segmentation.
• Current methods struggle with prediction errors from data uncertainty and loss instability from model uncertainty.

Purpose of the Study:

• To propose an Uncertainty-Guided Adaptive Correction (UGAC) framework to address limitations in semi-supervised medical image segmentation.
• To improve segmentation accuracy and robustness by managing data and model uncertainties.

Main Methods:

• Developed a dual-path uncertainty rectification mechanism using normalized entropy and confidence-weighted fusion.
• Implemented adversarial consistency constraints with spectral normalization for regularization.
• Introduced a frequency-aware segmentation backbone (Freqfusion module) for adaptive spectral decomposition.

Main Results:

• UGAC demonstrated superior performance on MM-WHS, BUSI, M&Ms, and PROMISE12 datasets.
• Achieved robust generalizability across CT, MRI, and ultrasound modalities.
• Showcased significantly lower computational complexity compared to baseline UNet.

Conclusions:

• The UGAC framework effectively overcomes challenges of data and model uncertainty in semi-supervised medical image segmentation.
• UGAC offers a robust, generalizable, and computationally efficient solution for medical image analysis.