A Unified Pipeline for Consistent Multi-Label Mask Generation in Pediatric Cardiac Segmentation

Abstract

Artificial Intelligence based cardiac image segmentation requires consistent labels, harmonized volumes, and computationally efficient dataset. The present study proposes a reproducible pipeline that (i) merges segmented cardiac structures into a single multi-label mask (0: background, 1: aorta, 2: chambers, 3: pulmonary structures), (ii) resamples masks to perfectly match scan dimensions using 3D Slicer's Resample Image (BRAINS) module (nearest-neighbor interpolation), and (iii) performs automatic region-of-interest (ROI) cropping by adding 10-pixel margin around mask-derived bounds. On our internal dataset, the workflow standardizes inputs, reduces data volume substantially, and preserves full cardiac anatomy. The resulting data enabled stable multi-class and task-specific training and contribution to a better Dice score in downstream models. The ROI script operate on a simple folder structure and can be executed batch-wise, supporting anonymity, interoperability, and rapid experimentation.