Deep learning model automates whole-heart volumetrics and hemodynamics from 4D flow CMR magnitude images

Deep learning on 4D flow cardiovascular magnetic resonance (CMR) magnitude images enabled automated whole-heart segmentation (Dice 0.88) and hemodynamic metrics (r≥0.88 vs manual). 4D flow magnitude had excellent ventricular volumes vs cine (ρ=0.98).

• 4D flow magnitude images correlated well with standard cine for left and right ventricular end-diastolic volumes (ρ=0.98 and 0.97, ICC≥0.98).
• Deep learning model achieved mean Dice similarity coefficient of 0.88 for whole-heart segmentation.
• AI-derived peak hemodynamic metrics strongly agreed with manual contours (r≥0.88).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.