What does this research mean for the field?

Deep learning automation of 4D flow CMR magnitude imaging provides accurate whole-heart volumetrics and haemodynamic assessments comparable to standard cine imaging and manual segmentation. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To assess the anatomical accuracy of 4D flow CMR magnitude images against standard cine imaging and validate a deep learning model for automated whole-heart segmentation and haemodynamics.

July 11, 2026European Heart Journal - Imaging Methods and PracticeOpen Access

Automated Whole-Heart Volumetrics and Haemodynamics from 4D Flow CMR Magnitude Images: Development and Validation of a Deep Learning Model

Q: What are the key findings of this study?

Automated deep learning segmentation of 4D flow CMR magnitude images achieved a mean Dice similarity coefficient of 0.88, providing accurate whole-heart volumetrics comparable to manual analysis.

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Cardiovascular disease requiring CMR (n=40). Intervention: Automated deep learning segmentation of 4D flow CMR magnitude images vs. Manual segmentation and standard cine SSFP imaging. Primary outcome: Mean Dice similarity coefficient (DSC) for automated segmentation of cardiac chambers and great vessels.

View Full Paper

Ask AI

Bookmark

Key result

Automated 4D flow CMR deep learning achieves high segmentation accuracy, with a mean 0.88 Dice score.

n=40

Why the study?

4D flow CMR provides comprehensive haemodynamic assessment but is limited by complex post-processing, prompting the need to validate 4D flow magnitude images against cine imaging and automate segmentation using AI.

Does an automated deep learning model accurately segment 4D flow CMR magnitude images for whole-heart volumetrics and haemodynamics compared to manual segmentation and standard cine imaging in patients undergoing CMR?

Population

40 patients prospectively identified from the PREFER-CMR registry

Comparison

4D flow magnitude imaging and AI segmentation vs standard cine imaging and manual segmentation

Design

Two-stage prospective registry-based validation and deep learning development study

Authors

AGAlexander GallUniversity of East Anglia CGCiaran Grafton‐ClarkeCardiac Imaging RLRui LiUniversity of East Anglia

Discussion

Member takes

Implication

May enable single-acquisition CMR volumetrics; hypothesis-generating before routine clinical adoption.

Key Points

To assess the anatomical accuracy of 4D flow CMR magnitude images against standard cine imaging and validate a deep learning model for automated whole-heart segmentation and haemodynamics.
Prospectively evaluated 40 patients (mean age 69.0 ± 17.2 years; 60.0% male) from the PREFER-CMR registry undergoing standard cine stacks and 4D flow CMR.
Stage 1 validated manual 4D flow magnitude segmentations against standard cine imaging for ventricular volumetrics and normative values for atrial volumetrics.
Stage 2 trained and validated a fully automated deep learning model (n=30 training, n=10 validation) and compared derived haemodynamic metrics against manual contours.
Manual 4D flow magnitude analysis strongly correlated with cine imaging for LVEDV (ρ=0.98, ICC=0.99, P<0.05) and RVEDV (ρ=0.97, ICC=0.98, P<0.05).
The automated deep learning model achieved an overall mean Dice similarity coefficient of 0.88, peaking at 0.92 for the aorta.
Haemodynamic metrics derived from automated versus manual contours demonstrated high agreement across all four chambers (r ≥ 0.88 for all peak metrics).

Study Design

Type

Observational (n=40)

Multicenter

Structured PICO

Population

40 adult patients with clinical indications for CMR, including heart failure and valvular disease, were evaluated to validate a deep learning model for automated whole-heart volumetrics.

Exposure

Automated 3D segmentation of the left ventricle, right ventricle, left atrium, right atrium, aorta, and main pulmonary artery on 4D flow MRI magnitude images using a 3D full-resolution nnU-Net deep learning model.

Comparator

Manual segmentation of 4D flow magnitude images and standard cine balanced steady-state free precession (bSSFP) imaging.

Outcome

Anatomical accuracy (volumetrics including EDV, ESV, SV, EF) and advanced haemodynamic metrics (kinetic energy, vorticity, viscous energy loss).surrogate

Automated deep learning segmentation of 4D flow CMR magnitude images provides accurate whole-heart volumetrics and haemodynamics, enabling rapid and comprehensive physiological assessment from a single acquisition.

Limitations

Single-centre study with a relatively small cohort of 40 patients.
Systematic bias observed in LVESV suggests further training with a larger, more diverse dataset is required.
All data were acquired on a single 1.5 T scanner from one vendor, requiring external validation across scanner vendors and field strengths.
Significant biases were observed for right ventricular energy loss and atrial parameters due to small differences in segmentation boundaries.
Small sample size
Single-center study
Lower spatial resolution and contrast-to-noise ratios of 4D flow magnitude images compared to optimized bSSFP cine sequences
Systematic bias observed in LVESV and atrial ESV with the AI model

Cite This Study

Gall et al. (2026) conducted an observational in Cardiovascular disease requiring CMR (n=40). Automated deep learning segmentation of 4D flow CMR magnitude images vs. Manual segmentation and standard cine SSFP imaging was evaluated on Mean Dice similarity coefficient (DSC) for automated segmentation of cardiac chambers and great vessels. Automated deep learning segmentation of 4D flow CMR magnitude images achieved a mean Dice similarity coefficient of 0.88, providing accurate whole-heart volumetrics comparable to manual analysis.

synapsesocial.com/papers/6a51e019c18d7f28ca500b4a https://doi.org/10.1093/ehjimp/qyag125

View Full Paper

Ask AI

Bookmark

Automated Whole-Heart Volumetrics and Haemodynamics from 4D Flow CMR Magnitude Images: Development and Validation of a Deep Learning Model

Discussion

Member takes

Implication

Key Points

Study Design

Structured PICO

Limitations

Cite This Study

Automated Whole-Heart Volumetrics and Haemodynamics from 4D Flow CMR Magnitude Images: Development and Validation of a Deep Learning Model

Discussion

Member takes

Implication

Key Points

Study Design

Structured PICO

Limitations

Cite This Study

Also Consider

Also Consider