Exploiting XAI maps to improve MS lesion segmentation and detection in MRI

Explainability can improve AI accuracy in medical imaging: By analyzing visual explanations (saliency maps) of what the AI model "sees" when making decisions, researchers improved the detection of multiple sclerosis lesions in brain MRI scans, increasing the model's precision from 62.65% to 78.17%.

Understanding false positives is clinically valuable: The study found meaningful differences in the explanation maps between correct detections and false alarms, suggesting that examining why an AI makes mistakes—not just that it makes them—can help refine its performance and reduce unnecessary patient concern from incorrect diagnoses.

Explainability methods have practical applications beyond transparency: Rather than serving only as a "black box" explanation tool, the researchers used explainability techniques as an active component to boost model performance, demonstrating that interpretability and accuracy can work together rather than as trade-offs.

Rigorous validation is essential for clinical AI claims: The study tested improvements on a separate patient population (37 test patients) with confidence intervals reported, following good scientific practice—an important standard when AI systems are proposed for real medical decision-making.

Combining AI approaches can enhance reliability: By layering a secondary machine learning model (logistic regression) on top of explainability maps from the primary model, the researchers achieved meaningful performance gains, suggesting that hybrid approaches may be more trustworthy than single AI systems alone.