Semi-Supervised Learning for Congenital Heart Disease Prediction in Fetal Echocardiography

Fetal echocardiography is widely used for prenatal congenital heart disease (CHD) screening, but automated analysis is difficult due to strong speckle noise, large view variation, and severe class imbalance. This challenge evaluates semisupervised anatomical segmentation and multi-label CHD classification, and the final ranking is a weighted combination of classification score, segmentation score, and processing time. We propose a two-stage pipeline. We first segment viewspecific cardiac structures to obtain an anatomical mask. We then perform multi-label CHD classification with a patch-andglobal framework using a self-supervised Vision Transformer backbone. The predicted mask guides region-of-interest patch sampling and provides weak conditioning features, while view-aware constraints suppress view-incompatible labels to reduce false positives. Experiments are conducted on 2,500 ultrasound slices across four standard views, with 391 slices having pixel-wise anatomical annotations and case-level labels for seven CHD types. Our method achieves an overall score of 44.14 with an F1-score of 40.77, a DSC of 77.72, and an NSD of 61.45, outperforming the official baseline (35.89 overall, 34.20 F1, 65.48 DSC, 45.55 NSD). Code is available at https://github.com/hwu918945-alt/isbi-2026-FETUS-Semi-Seg-CHDClassification.