JMIR Publications

ABSTRACT

Background:

Cystic hygroma is a high-risk prenatal ultrasound finding that portends high rates of chromosomal abnormalities, structural malformations, and adverse pregnancy outcomes. Automated detection can increase reproducibility and support scalable early screening programs, but supervised deep learning methods are limited by small labelled datasets.

Objective:

This study assesses whether ultrasound-specific self-supervised pretraining can facilitate accurate, robust deep learning detection of cystic hygroma in first-trimester ultrasound images.

Methods:

We fine-tuned the Ultrasound Self-Supervised Foundation Model with Masked Autoencoding (USF-MAE), pretrained on over 370,000 unlabelled ultrasound images, for binary classification of normal controls and cystic hygroma cases used in this study. Data were accessed for this research between 09/12/2025 and 31/12/2025. Performance was evaluated on the same curated ultrasound dataset, preprocessing pipeline, and 4-fold cross-validation protocol as for the DenseNet-169 baseline, using accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (ROC-AUC). Model interpretability was analyzed qualitatively using Score-CAM visualizations.

Results:

USF-MAE outperformed the DenseNet-169 baseline on all evaluation metrics yielding a mean accuracy of 0.96±0.02, sensitivity of 0.94±0.06, specificity of 0.98±0.02, and ROC-AUC of 0.98±0.02 compared to 0.93±0.03, 0.92±0.07, 0.94±0.01, and 0.94±0.03 for the DenseNet-169 baseline, respectively. Qualitative Score-CAM visualizations of model predictions demonstrated clinical relevance by highlighting expected regions in the fetal neck for both positive and negative cases.

Conclusions:

Ultrasound-specific self-supervised pretraining enables deep learning models to detect cystic hygroma with higher accuracy and robustness than supervised convolutional neural networks trained from scratch. This data-efficient approach requires minimal manual annotation and demonstrates the potential of self-supervised learning for scalable decision support in early prenatal anomaly screening.