Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Decipher-MR: A Foundation Model That Could Accelerate MRI-Based Diagnostic AI
Decipher-MR represents a significant advance in medical imaging AI, offering a reusable foundation model trained on 200,000 diverse MRI scans that consistently outperforms existing approaches across disease detection, anatomy localization, and cross-modal retrieval tasks.
Background
Medical imaging AI has historically struggled with poor generalization—models trained on specific datasets or pathologies often fail when applied to new contexts. Decipher-MR tackles this challenge by creating a 3D MRI-specific vision-language foundation model. Trained on over 22,000 studies using self-supervised learning combined with radiologist report guidance, the model learns generalizable MRI representations across diverse anatomical regions, imaging sequences, and pathologies.
Key Findings
Decipher-MR demonstrated consistent improvements over existing foundation models and traditional task-specific approaches across multiple benchmarks. Its modular design enables efficient deployment: lightweight task-specific decoders attach to a frozen pretrained encoder, eliminating the need to retrain the entire model for new clinical applications. The model validated successfully across multiple public datasets (ADNI, PI-CAI, ACDC, LLD-MMRI, MRART, AMOS).
Why It Matters
This work addresses critical barriers in medical AI: data scarcity and poor cross-domain generalization. By providing a reusable foundation trained on heterogeneous MRI data, Decipher-MR could accelerate development of AI diagnostic tools across diverse clinical applications while reducing dependence on large labeled datasets for specialized domains.
Limitations
Reproducibility details remain limited. Real-world clinical deployment testing with varied scanner types and acquisition protocols is still needed to fully validate generalization claims.
Original paper: Decipher-MR: a vision-language foundation model for 3D MRI representations. — NPJ digital medicine. 10.1038/s41746-026-02596-4
