Vision-language framework for multi-sequence brain magnetic resonance imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Structural magnetic resonance imaging (MRI) is crucial for diagnosing neurological disorders.
• Automated interpretation of multi-sequence brain MRI faces challenges due to cross-sequence reasoning and protocol variability.

Purpose of the Study:

• To introduce ReMIND, a vision-language modeling framework for comprehensive multi-sequence and multi-volumetric brain MRI analysis.
• To address limitations in automated brain MRI interpretation.

Main Methods:

• ReMIND was trained on over 73,000 deidentified patient visits (850,000+ MRI sequences) with radiology reports.
• The framework utilized large-scale instruction tuning on over one million clinical question-answer pairs and supervised fine-tuning for report generation.
• Modality-aware reranking and correction strategies were employed during inference for report generation.

Main Results:

• ReMIND demonstrated cross-cohort generalization on independent external datasets.
• The model suppressed unsupported modality claims while maintaining linguistic fluency and clinical coherence.
• The framework advanced consistent and equitable brain MRI interpretation.

Conclusions:

• ReMIND represents a significant advancement in automated brain MRI analysis.
• Prospective evaluation is warranted to support diagnosis and management of neurological conditions.
• The findings support the potential for more equitable and consistent neurological disorder diagnosis.