Conservative AI for Medical Image Restoration in CT-CTA

A recent study introduces a conservative AI framework aimed at enhancing medical image restoration, focusing on intracranial CT and CT angiography (CTA) scans. This framework seeks to elevate image quality while avoiding unregulated alterations in clinically significant areas, including small vessels and cues related to aneurysms near high-contrast anatomical boundaries. Utilizing a residual-bounded 2.5D restoration model, it is trained on synthetically degraded CT/CTA images, incorporating a learned residual to the original center slice via an edit-control map that restricts the extent and magnitude of changes. The evaluation process involves an aneurysm-relevant image-recovery matrix, comparisons with a Gaussian baseline, Monte Carlo stability assessments, anatomical localization of significant edits, and external validation on low-dose CT. The paper can be found on arXiv with the identifier 2605.16458.