Self-Supervised Text-Vision Alignment for Automated Brain MRI Abnormality Detection: A Multicenter Study (ALIGN...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Medical Informatics
• Machine Learning for Healthcare

Background:

• Expert-labeled datasets are crucial but costly and time-consuming for training AI models in medical imaging.
• Leveraging existing free-text radiology reports offers a scalable alternative for AI model development.
• Brain MRI interpretation requires specialized knowledge, making automated detection highly valuable.

Purpose of the Study:

• To develop and validate a self-supervised text-vision framework for detecting brain MRI abnormalities.
• To eliminate the dependency on expert-annotated datasets by utilizing radiology reports.
• To assess the framework's diagnostic performance and generalizability across multiple institutions.

Main Methods:

• A retrospective and prospective multicenter study involving 81,936 brain MRI examinations and reports for training/internal testing.
• Development of a neuroradiology language model (NeuroBERT) using self-supervised learning for report embeddings.
• Training convolutional neural networks to map MRI scans to embeddings, followed by text-image similarity scoring for abnormality detection.

Main Results:

• The framework achieved an area under the receiver operating characteristic curve (AUC) of 0.95 for normal vs. abnormal classification.
• Excellent generalizability to external sites with examination-level AUCs ranging from 0.85 to 0.90.
• High performance in zero-shot classification tasks (mean AUC 0.89) and visual-semantic image retrieval (mean precision 0.84).

Conclusions:

• The developed self-supervised text-vision framework accurately detects brain MRI abnormalities.
• This approach successfully leverages free-text radiology reports, negating the need for expert-labeled training data.
• The framework demonstrates robust diagnostic capabilities and broad applicability in clinical settings.