Automated whole-heart volumetrics and haemodynamics from 4D flow CMR magnitude images: development and validation of a deep learning model.
Authors
Affiliations (4)
Affiliations (4)
- Norwich Medical School, University of East Anglia, Norwich Research Park, Norwich NR4 7UQ, UK.
- Department of Cardiology, Norfolk and Norwich University Teaching Hospitals, Norwich NR4 7UY, UK.
- Leeds Institute of Cardiovascular and Metabolic Medicine, University of Leeds, Leeds, UK.
- Division of Image Processing, Department of Radiology, Leiden University Medical Centre, Leiden, The Netherlands.
Abstract
4D flow cardiovascular magnetic resonance (CMR) offers a comprehensive haemodynamic assessment but is often limited by long acquisition times and complex post-processing. The magnitude images derived from 4D flow sequences contain time-resolved 3D anatomical information. We aimed to validate the anatomical accuracy of these images against standard cine imaging and develop an artificial intelligence (AI) model for automated segmentation to facilitate analysis. Forty patients prospectively identified from the PREFER-CMR registry underwent CMR, including standard cine stacks and 4D flow. The study consisted of two stages. In Stage 1, manual segmentation of the cardiac chambers and great vessels was performed on 4D flow magnitude images. These were validated against standard cine volumetrics (LV/RV) and normative reference values (LA/RA). In Stage 2, a fully automated deep learning algorithm was trained and validated. Advanced haemodynamic metrics were derived using both manual and AI segmentations to assess agreement. The study cohort (<i>n</i> = 40) had a mean age of 69.0 ± 17.2 years, and 60.0% were male. In Stage 1, 4D flow magnitude analysis demonstrated excellent correlations with cine measurements for LV end-diastolic volume (ρ = 0.98, ICC = 0.99) and RV end-diastolic volume (ρ = 0.97, ICC = 0.98). In Stage 2, the AI model achieved excellent segmentation performance (mean Dice similarity coefficient 0.88). Comparisons of haemodynamic metrics derived from AI vs. manual contours showed strong agreement (<i>r</i> ≥ 0.88 for all peak metrics). 4D flow magnitude imaging provides accurate volumetrics. Deep learning automation of this process is feasible, allowing for rapid, comprehensive assessment of cardiac structure, function, and advanced energetics.