Enhancing Brain MRI Anomaly Detection and Reasoning with ROI Rethink and Synthetic Data

Researchers introduce BrReMark, a framework that enhances brain MRI diagnosis by requiring AI models to explicitly mark and verify abnormal regions before reaching conclusions. The approach dramatically improves diagnostic accuracy and reduces false positives by 45.7% on out-of-distribution data, addressing critical trust and hallucination issues in medical AI systems.

BrReMark tackles a fundamental problem in medical AI: models that generate diagnoses without showing their reasoning or spatial grounding. This black-box behavior creates clinical liability since radiologists cannot audit why a model reached its conclusion, and systems frequently hallucinate findings on normal scans. The framework enforces interpretability by requiring models to generate hypotheses about abnormalities, mark their locations with bounding boxes, and re-examine marked regions to verify conclusions—mimicking the clinical reasoning process itself.

The approach emerges from broader recognition that medical AI systems need explainability and robustness to be clinically viable. Previous models achieved high accuracy metrics but failed at generalization and failed to reduce false positives, particularly on rare or novel pathologies. BrReMark combines supervised fine-tuning on structured reasoning trajectories with reinforcement learning that rewards both localization accuracy and diagnostic reasoning quality. The addition of synthetic data augmentation through domain randomization further improves out-of-distribution performance, addressing the critical gap between research benchmarks and real-world clinical variability.

For healthcare institutions and medical AI developers, this represents a significant step toward trustworthy deployment. The 45.7% reduction in false positives directly impacts clinical workflows by reducing unnecessary follow-up imaging and specialist consultations. The framework's strong performance on out-of-distribution NOVA benchmarks suggests models can generalize beyond training data—essential for handling patient populations and pathologies not well-represented in training sets.

Future work likely focuses on extending this hypothesis-verification framework to other medical imaging modalities and integrating explainable AI into regulatory compliance frameworks for medical device approval.

• →BrReMark enforces AI interpretability in medical diagnosis by requiring explicit region marking and verification before conclusions.
• →The framework achieves 45.7% reduction in false positives on out-of-distribution data, addressing hallucination problems critical for clinical adoption.
• →Combining reinforcement learning with synthetic pathology augmentation improves generalization to rare and novel brain abnormalities.
• →Diagnostic accuracy reaches 45.26% with explicit grounding, significantly outperforming ungrounded baselines on structured reasoning tasks.
• →The approach establishes a practical path toward trustworthy AI deployment in medical imaging through enforced hypothesis-verification workflows.