From Precision to Planetary Health: the Carbon Cost of AI in Multimodality Cardiovascular Imaging

Artificial intelligence is becoming a defining component of multimodality cardiovascular imaging, with applications spanning acquisition, reconstruction, post processing, quantification, reporting, and clinical decision support. Yet the environmental cost of this transformation remains largely absent from scientific reporting. This editorial argues that, in a health sector already under pressure to decarbonise, the ecological footprint of artificial intelligence can no longer be considered negligible. Cardiovascular imaging is intrinsically energy intensive, and artificial intelligence may add substantial computational demand, particularly during model development, hyperparameter tuning, and repeated training. The key asymmetry is that training often consumes far more energy than inference, while current publications usually report performance without disclosing energy use, GPU hours, or associated carbon emissions. We propose that efficiency should become a core methodological metric alongside discrimination, calibration, and clinical utility. A pragmatic framework for carbon transparency is outlined based on four steps: measuring computation related energy consumption, accounting for data centre overhead, converting electricity use into carbon emissions according to grid intensity, and reporting the full development footprint rather than only the final selected model. Such reporting could be summarised through a simple carbon label including emissions per training run, cumulative development emissions, and emissions per patient or per 1000 inferences. It calls on journals and reviewers to request carbon transparency statements. Far from opposing innovation, this perspective supports a more frugal, reproducible, equitable, and responsible form of medical artificial intelligence, while also recognising that appropriately deployed artificial intelligence may help reduce the overall environmental burden of cardiovascular imaging workflows.