Uncertainty-Aware Super-Resolution for Mammography Phantoms using a Dropout-Enabled SwinIR.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Deep learning models like SwinIR lack transparency, hindering clinical trust in mammography.
• Uncertainty quantification is crucial for assessing the reliability of AI-driven medical imaging.

Purpose of the Study:

• To develop and evaluate a framework for integrating uncertainty quantification into the SwinIR super-resolution model for mammography.
• To address the "black box" limitation and enhance the interpretability of AI models in medical imaging.

Main Methods:

• Incorporated Monte Carlo (MC) Dropout into the SwinIR model to generate super-resolved mammography images and pixel-wise uncertainty maps.
• Evaluated the framework on a standardized mammography phantom using standard super-resolution and image enhancement tasks.

Main Results:

• The L1 SwinIR model achieved a high structural similarity index (0.982) in super-resolution, outperforming bicubic interpolation.
• Uncertainty maps demonstrated context-dependent behavior, with low uncertainty for preserved high-frequency structures and high uncertainty for sharpened features.
• Analysis revealed an overconfident rate in low-contrast regions, where low uncertainty did not always correlate with low error.

Conclusions:

• The developed framework provides a transparency tool for AI in mammography, with uncertainty reflecting model behavior rather than direct error.
• Findings suggest potential for improved interpretability in medical imaging, but further validation on clinical data is necessary.
• Publicly available source code facilitates further research and development in AI-driven medical image analysis.