Vision-Language Models Act as Semantic Regularizers to Enhance MRI Reconstruction

maier_ak · x · 2026-08-13

Researchers introduced a semantic regularizer for MRI reconstruction. This regularizer plugs into any existing deep-learning MRI pipeline as a contrastive loss. By utilizing a pre-trained vision-language foundation model, it pulls the semantic representation of the under-sampled reconstruction toward the target distribution derived from ground-truth or a text prompt, teaching the algorithm what a "good" image looks like visually and conceptually.

Original post →

Vision-Language Models Act as Semantic Regularizers to Enhance MRI Reconstruction

More from Research