Automated reconstruction of dynamic 2D fetal cardiac MRI using deep learning in late third-trimester fetuses.

INTRODUCTION Fetal cardiovascular magnetic resonance (CMR) suffers from motion corruption and cardiac gating challenges. Data-driven motion correction and gating remain dependent on manual region-of-interest (ROI) selection, limiting clinical utility through resource demands, processing time, and inter-observer variability. We developed an nnU-Net model to automatically select ROIs and integrated it into a cine reconstruction pipeline.

Methods

2327 real-time images acquired as multi-slice stacks from 23 late third-trimester pregnancies (34-36 weeks gestational age), comprising 21 healthy fetuses and 2 with congenital heart disease, were used to train and test the model. Images varied in orientation, temporal resolution, and signal-to-noise ratio. Manual and automatic ROI selections were compared using the Dice similarity coefficient, ROI sizes, and ROI centroids, and were separately input into a cine reconstruction pipeline with motion correction and metric-optimized gating. Cines were compared between methods using in-plane translational parameters, RR intervals, mutual information (MI), the blind Perception-derived Image Quality Evaluator (PIQE), and qualitative anatomical review.

Results

The nnU-Net model achieved a mean Dice score of 0.83 ± 0.23 with 81% of scores exceeding 0.8. No significant differences were found between ROI sizes (p = 0.13), translational parameters (along x-dimension: p = 0.58 or y-dimension: p = 0.11), or MI values (p = 0.59). Automatic and manual cines showed broadly comparable image quality, with statistically lower PIQE scores for automatic cines (p < 0.001) but a small absolute difference (∼0.5 points) of uncertain clinical significance. Cines displayed qualitatively comparable anatomical fidelity.

Conclusions

Deep-learning-automated ROI selection for fetal CMR reconstruction is feasible and addresses a key barrier to clinical adoption of the modality.