3D U-Net-Based Deep Learning Approach for Pediatric Cardiovascular MRI Segmentation

Abstract

Accurate segmentation of cardiovascular structures in pediatric patients is essential for clinical evaluation, but variability in imaging data and the labor-intensive nature of manual segmentation remain major challenges. The given paper provides a reproducible deep-learning model intended to counteract these issues. With our deep learning approach, cardiac MRI data is standardized by (i) converting anatomical labels into a four-class mask (involving background, aorta, chambers, pulmonary structures), (ii) aligning images and masks on coordinates using nearest-neighbor resampling, and (iii) automatically cropping the region-of-interest (ROI). To handle class imbalance, we trained a 3D U-Net with the Focal Tversky Loss function, using both public and private clinical datasets. To improve robustness and reduce outlier errors, we applied model ensembling by averaging predictions from independently trained networks. By automating preprocessing and segmentation, our approach provides a pediatric cardiac analysis tool that reduces manual workload and increases reproducibility.