Deep Learning Analysis of Prenatal Ultrasound for Identification of Ventriculomegaly

To develop and evaluate a deep learning model capable of detecting ventriculomegaly on prenatal ultrasound images using a foundation model pre-trained specifically on ultrasound data.

A vision transformer-based ultrasound self-supervised foundation model with masked autoencoding (USF-MAE) was fine-tuned for binary classification of fetal brain ultrasound images as normal or ventriculomegaly. The encoder had been pre-trained on more than 370,000 ultrasound images from the OpenUS-46 corpus. For this study, it was adapted and fine-tuned on a curated data set of fetal brain images. Model performance was evaluated through fivefold cross-validation as well as an independent test cohort. Accuracy, precision, recall, specificity, F₁-score and area under the receiver operating characteristic and precision-recall curves were reported as the performance metrics. Eigen-CAM and Grad-CAM were used to visualize model attention.

USF-MAE reported an F₁-score of 91.76% for the cross-validation set and 91.78% for the test set. The model outperformed all the baseline models, which included VGG-19, ResNet-50, ViT-B/16 and MoCo v3. The model reported a mean test precision of 94.47% and an accuracy of 97.24%. Activation maps indicated that the model consistently focused on the ventricular region when identifying ventriculomegaly.

Pre-training on a large, ultrasound-specific corpus improved classification performance and generalization for ventriculomegaly detection. The USF-MAE framework provided strong accuracy, reliability and explainability, showing potential as a robust tool for the assessment of fetal brain structures on prenatal ultrasound.