MARS foundation model tops 41 of 44 multi-sequence MRI benchmarks

MARS, a foundation model pretrained on 336,476 multi-sequence MRI scans, ranked first in 41 of 44 downstream tasks (diagnosis, segmentation, etc.) with significant gains, showing generalizability across anatomy and sequences.

• Pretrained on 336,476 volumetric scans from 34 datasets covering 10 anatomical structures and multiple MRI sequences.
• Benchmark comprised 44 tasks: diagnosis, segmentation, registration, progression prediction, and report generation.
• Strong performance on heterogeneous and external datasets suggests robust generalization.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.