A self-supervised learning (SSL) framework trained on over 48,000 cardiac MRI images achieved a mean absolute error of 3.70% for intra-cardiac blood oxygen saturation — a 15% improvement over supervised baselines and traditional radiomics approaches. The system pre-trained ResNet and vision transformer encoders using contrastive learning (SimCLR) and masked image modeling on unlabeled cine and T2 oximetry CMRI data, then fine-tuned for oxygen regression with built-in uncertainty quantification.
Currently, accurate intra-cardiac oxygen saturation requires right-heart catheterization — an invasive procedure carrying procedural risks including arrhythmia, vessel injury, and infection. Non-invasive alternatives have long been sought but stalled by the annotation bottleneck: deep learning models need large labeled datasets, which are expensive and scarce in specialized cardiac imaging. This SSL approach sidesteps that constraint by learning generalizable representations from unlabeled data — a strategy that has transformed natural language processing and is now reaching cardiovascular diagnostics.
A 3.70% MAE is clinically meaningful: values above roughly 5% can shift clinical classification of shunt severity or pulmonary hypertension. However, critical caveats apply. Validation appears limited to a single institutional dataset; external generalizability across scanner vendors and patient populations remains untested. The regression target is continuous oxygen saturation, not clinical decision endpoints. As a preprint posted on medRxiv and not yet peer-reviewed, these results should be treated as preliminary — independent replication and prospective clinical validation are necessary before any diagnostic deployment.