Explainable automated image quality assessment for magnetic resonance imaging through prediction of defect maps

Motivation: Monitoring image quality is crucial for magnetic resonance imaging. However, it can present a challenge for inexperienced technologists or in high-volume scenarios. Goal(s): Develop an automated method for image quality monitoring to assist the technologists. Approach: Train Swin-U-net model to predict defect maps using corrupted image as training input and difference map between the original and corrupted image as label. Results: On evaluating with clinical annotations, the proposed approach resulted in accurate image quality predictions. The predicted defect maps provide a novel explainable image highlighting affected areas of the input image and offering meaningful visual explanations for the technologist regarding the quality judgement. Impact: Clinical adoption of automated image quality assessment methods can help technologists in automatically monitoring image quality. Meaningful visual explanations offered by the proposed approach could help in building trust in the method and allow fast false positive detection by technologists.