Structure-Aware V-Net Framework for 3D Multi-Class Segmentation of Pediatric Cardiovascular Structures

Abstract

This research shows a structure-aware deep learning model. The model segments 3D multi-class pediatric cardiovascular anatomy, focusing on small complex structures, like the pulmonary arteries, also it is based upon a modified V-Net architecture. The multi-center dataset includes expert-annotated pediatric Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) volumes. The model underwent training through utilization of this specific dataset. For resolving class disequilibrium in conjunction with morphological variance, investigators incorporated a supplementary decoder division for specialization within pulmonary artery segmentation. The pipeline represents standardized preprocessing measures like resampling, normalization, also cropping, furthermore it represents advanced augmentation techniques. Evaluation was performed using Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) metrics. This assessment revealed that the partitioning of minute anatomy progressed noticeably. The proposed architecture shows support is offered for preoperative planning in congenital heart disease (CHD) cases because structure-specific modifications are able to improve segmentation accuracy plus consistency in pediatric cardiovascular imaging.