Large-scale multi-sequence pretraining for generalizable MRI analysis in versatile clinical applications.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Multi-sequence MRI is crucial for clinical diagnosis but faces challenges with data heterogeneity.
• Heterogeneity limits the generalizability and clinical translation of deep learning models in medical imaging.

Purpose of the Study:

• To develop a large-scale MRI foundation model (MARS) to address heterogeneity in multi-sequence MRI data.
• To create robust and generalizable representations for diverse clinical applications using a novel pretraining strategy.

Main Methods:

• Collected 64 datasets across 10 anatomical structures and multiple MRI sequences.
• Curated 336,476 volumetric scans from 34 datasets to build a multi-organ, multi-sequence MRI pretraining corpus.
• Developed a novel pretraining strategy to disentangle anatomy-invariant features from sequence-specific variations.

Main Results:

• Established a benchmark of 44 downstream tasks including diagnosis, segmentation, registration, and report generation.
• MARS achieved first-place rankings in 41 out of 44 benchmarks with statistically significant improvements.
• Demonstrated strong performance on heterogeneous and external datasets, highlighting its generalizability.

Conclusions:

• MARS serves as a scalable foundation model for versatile multi-sequence MRI analysis.
• The novel pretraining strategy effectively handles MRI data heterogeneity, improving deep learning model performance.
• MARS shows significant potential for advancing clinical translation of AI in medical imaging.