3D Cardiac MRI Segmentation Using Spatio-Temporal Convolutional Neural Networks with Shape-Aware Loss Functions

Abstract
This research proposes a high-precision spatio-temporal deep learning framework for 3D cardiac MRI segmentation, integrating dynamic feature encoding with anatomically guided supervision. Leveraging a hybrid architecture based on 3D convolutional layers and ConvLSTM modules, the model captures both spatial structures and temporal motion patterns across full cardiac cycles. Shape-aware loss functions—including signed distance function (SDF) loss and boundary-based constraints—are incorporated to ensure anatomical fidelity, reducing segmentation artifacts and improving topological correctness. Evaluated across benchmark datasets such as ACDC, M&Ms, and UK Biobank, the model consistently achieved Dice scores above 92.1% and reduced Hausdorff distances to 5.2mm. Compared to conventional 2D/3D CNNs, this framework showed a 3.7–4.5% increase in segmentation accuracy and an 18–25% reduction in boundary leakage, particularly in complex regions such as the apex and basal slices. The suggested technique accomodates strong generalization over pathological cases, ensuring temporal coherence and clinical applicability. Its encoder-decoder framework is still flexible to different sequence lengths and resolutions to facilitate scalable deployment in real-time diagnostic and surgical planning applications.

Author
Sazan Kamal Sulaiman

DOI
https://doi.org/10.1109/ICCR64138.2025.11292565

ISSN
2836-8142

Publish Date: 19-Dec-2025