Single-Subject Multi-View MRI Super-Resolution via Implicit Neural Representations
New framework eliminates need for large training datasets by learning directly from a patient's own multi-view scans.
A research team from Cornell Tech and Erasmus MC has introduced SIMS-MRI (Single-Subject Implicit Multi-View Super-Resolution for MRI), a breakthrough AI framework that addresses a critical bottleneck in clinical MRI. Traditional MRI often acquires fast, anisotropic volumes with high in-plane but low through-plane resolution, requiring multiple orientations for complete anatomical information. Conventional integration methods rely on registration and interpolation, which can degrade fine structural details, while recent deep learning approaches depend on large-scale training datasets that introduce cohort-level biases and reduce clinical reliability.
SIMS-MRI innovates by combining a multi-resolution hash-encoded implicit neural representation with learned inter-view alignment. This allows the system to generate a spatially consistent, isotropic reconstruction using only the anisotropic multi-view scans from a single patient, eliminating the need for any pre- or post-processing. The framework is fundamentally self-supervised, learning directly from the target patient's scans, which removes dependency on external training data and improves generalizability to individual clinical cases. The team validated SIMS-MRI on both simulated brain MRI and clinical prostate MRI datasets, demonstrating its practical utility. The code will be made publicly available, promoting reproducibility and further development in the medical AI community.
- Eliminates need for large training datasets by using self-supervised learning on single-patient data
- Combines implicit neural representations with multi-resolution hash encoding for efficient 3D reconstruction
- Validated on clinical prostate MRI data, showing direct applicability to real-world medical imaging
Why It Matters
Enables more reliable, patient-specific medical imaging without the biases and data requirements of traditional AI models.