MRI field-strength enhancement aims to give low-field scanners the image quality of high-field systems, but existing methods often hallucinate anatomy because they process all brain regions uniformly and rely on scarce paired data. UniField++ addresses this with two cooperating agents: a vision-language-model planner that predicts which regions are risky and formulates targeted restoration plans, and a flow-matching enhancer that follows those plans while looping back over high-risk areas for extra refinement.
The framework is trained across multiple field-strength transitions—including 64mT-to-3T and 3T-to-7T—so both agents learn field-dependent tissue contrast and anatomical visibility patterns that generalize beyond a single transition. Radiologists' annotations of hallucination-prone regions and correction plans supervise the planner, and the enhancer uses loop reasoning along the flow-matching trajectory to repeatedly process difficult areas.
Across the evaluated tasks, UniField++ reports average improvements of 2.74 dB in PSNR and 14.05 percentage points in SSIM over comparison methods, with corresponding gains of 3.97 dB and 13.92 percentage points over its ablation baseline. It also maintains quality on an external cohort with an unseen field strength and improves downstream whole-brain segmentation Macro Dice by an average of 0.1457 over the original low-field images. As a single source, the paper presents no conflicting findings; its claims are limited to the reported experiments and the authors' own comparisons.