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

This study was aimed at presenting a framework integrating uncertainty quantification into the SwinIR super-resolution model for mammography, addressing the "black box" limitation that hinders clinical trust. Monte Carlo (MC) Dropout was incorporated into SwinIR to generate both high-resolution images and corresponding pixel-wise uncertainty maps. The framework was systematically evaluated on a standardized mammography phantom in two scenarios: a standard 4 × super-resolution task and an enhancement task applied to the original high-resolution image. In the standard task, the L1 model achieved a structural similarity index of 0.982, outperforming the bicubic interpolation baseline (0.978). The uncertainty maps revealed context-dependent behavior: low uncertainty for high-frequency structures destroyed by downsampling, high uncertainty when sharpening identifiable structures in the original image, and consistently low uncertainty for low-contrast objects. Quantitative analysis showed a near-zero correlation between absolute error and uncertainty in low-contrast regions, indicating an overconfident rate where approximately 0.3% of pixels showed high error but low uncertainty. High uncertainty does not indicate error but reflects active intervention on recognizable structures, whereas low uncertainty may represent either true stability or inaction in the face of insufficient evidence. In this phantom study, uncertainty served as a transparency tool rather than a measure of correctness. As a proof-of-concept, these findings suggest potential for enhancing interpretability, though further validation on clinical data is required to establish utility and safety in patient imaging. Source code is publicly available at https://github.com/evh-5150/DICOM_SR_Uncertainty .