Deep Learning Automated Measurement of Shunt Severity with Estimation of Uncertainty in 4D Flow MRI.

Purpose To assess the feasibility of a deep learning (DL) system to fully automate systemic and pulmonary blood flow measurement with four-dimensional (4D) flow MRI. Materials and Methods A total of 188 clinical 4D flow MRI examinations were retrospectively collected to develop a DL system that serially performs (a) three-dimensional localization and (b) two-dimensional segmentation to compute net aortic and pulmonary artery flow. For each vessel, the system computes the means and SDs (σ) across multiple locations. In a separate composite population of 71 patients with and without shunts, automated and manual measurements from trained physicians were compared with ground truth measurements from a senior radiologist using Pearson correlation and Bland-Altman analysis. Results Automated aortic and pulmonary artery flow estimates and shunt fraction calculations were performed in all 71 patients. Correlation with ground truth was strong (ρ = 0.832, 0.863, and 0.861, respectively; P < .001) with low mean bias (0.40 L/min, 0.20 L/min, and -0.10). High certainty (σ < 0.5 L/min) was observed in 83% (59 of 71) of aortic, 75% (53 of 71) of pulmonary, and 65% (46 of 71) of both aortic and pulmonary flow measurements. Within these subsets, correlations improved (ρ = 0.881, 0.989, and 0.948; P < .001) and closely matched expert readers. Conclusion A DL system can automatically measure blood flow from 4D flow MRI with accuracy comparable to that of trained physicians. Keywords: Deep Learning, Congenital Heart Disease, Shunt Severity, 4D Flow MRI, Convolutional Neural Networks Supplemental material is available for this article. © RSNA, 2026.