Improved cystic hygroma detection from prenatal imaging using ultrasound-specific self-supervised representation learning

TL;DR

This work addresses the challenge of detecting cystic hygroma in first-trimester fetal ultrasound images under limited labelled data. It leverages an ultrasound-specific self-supervised foundation model with Masked Autoencoding (USF-MAE), pretrained on a large unlabeled ultrasound corpus, and fine-tunes it for binary classification against a DenseNet-169 baseline trained from scratch. USF-MAE achieves higher mean accuracy ($0.96$), sensitivity ($0.94$), specificity ($0.98$), and ROC-AUC ($0.98$) than the baseline (e.g., $0.93$, $0.92$, $0.94$, $0.94$ respectively) with statistically significant improvement ($p = 0.0057$). The approach highlights data efficiency, robustness across cross-validation folds, and anatomically plausible interpretability via Score-CAM, supporting its potential as a scalable decision-support tool for early prenatal anomaly screening. The findings underscore the value of modality-specific self-supervised pretraining in prenatal imaging where labelled data are scarce, enabling more reliable automated screening and reducing reliance on expert annotations.

Abstract

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. This study assesses whether ultrasound-specific self-supervised pretraining can facilitate accurate, robust deep learning detection of cystic hygroma in first-trimester ultrasound images. 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. 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. USF-MAE outperformed the DenseNet-169 baseline on all evaluation metrics. The proposed model yielded a mean accuracy of 0.96, sensitivity of 0.94, specificity of 0.98, and ROC-AUC of 0.98 compared to 0.93, 0.92, 0.94, and 0.94 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. Paired statistical analysis using a Wilcoxon signed-rank test confirmed that performance improvements achieved by USF-MAE were statistically significant (p = 0.0057).

Figures (6)

• Figure 1: Comparison of first-trimester fetal ultrasound images. A) Normal fetus with typical NT thickness. B) Fetus with cystic hygroma, showing a markedly enlarged, multiloculated nuchal fluid collection consistent with lymphatic malformation.
• Figure 2: (A) Three-dimensional representation of the permitted grayscale area in HSV space according to the hue, saturation, and value thresholds. (B) The original ultrasound image with colored annotations. (C) The same image is seen in HSV space, with pixels outside the grayscale area highlighted to demonstrate how the preprocessing technique removes annotation artifacts. adapted from b5 by Walker et al.
• Figure 3: (A) The initial fetal ultrasound image. (B) Non-grayscale artifact locations are identified via a binary mask. (C) The image was processed using a Navier-Stokes-based b21 inpainting technique after artifact removal.
• Figure 4: Overview of the cystic hygroma classification workflow. The top panel shows the existing USF-MAE pretraining framework previously developed by our group b1, where a large, consolidated ultrasound dataset is preprocessed and employed for SSL using MAE. Randomly sampled image patches are hidden and reconstructed during pretraining, resulting in the encoder learning ultrasound-specific feature representations. The bottom panel depicts the contribution of this study, which extends this work by adapting the pretrained encoder for cystic hygroma classification, where the MAE decoder is discarded, and a classification head is added. The model is then fine-tuned using labelled fetal neck ultrasound data for diagnostic prediction support.
• Figure 5: ROC and precision-recall (PR) curves across cross-validation folds. (A) ROC curves for each of the four cross-validation folds, along with the mean ROC curve (solid blue line) and its variability. The dashed diagonal line represents the performance of a random classifier. The proposed model achieves a high mean ROC-AUC across folds, indicating strong discriminative performance. (B) Precision-recall curves across folds with the mean PR curve shown in blue. Iso-F1 score contours are overlaid to illustrate the trade-off between precision and recall. The consistently high PR-AUC indicates robust performance under class-imbalanced conditions.