Assessing Deep Learning's Ability to Segment Aortic Valve Calcifications in Contrast-Free Cardiac MRI

Motivation: Contrast-enhanced CT is the gold standard for TAVI planning but unsuitable for patients with kidney disease or contrast allergies. Contrast-free MRI is an alternative but lacks visualization of calcifications. Goal(s): Use Deep Learning models to segment calcified regions on contrast-free CMR Quantify from the CMR the calcified area, indicator of aortic stenosis severity Approach: We performed CTA-CMR registration with Elastix using aorta masks, then trained several SOTA Deeo Leaning models using CMR images and CTA-to-CMR registered calcification masks. Results: Segmentation of aorta valve calcifications showed low DSC (0.309), while aorta wall calcifications could not be estimated. Valves' calcified area was underestimated by 25.96%. Impact: Calcifications are not visible on contrast-free CMR to the human eye, so we tested AI segmentation using CTA-registered labels, achieving low results (DSC 0.309). While a larger dataset or improved registration may help, the task's physical feasibility remains uncertain.