Deep Learning Analysis of Prenatal Ultrasound for Identification of Ventriculomegaly

TL;DR

The paper tackles automated ventriculomegaly detection in prenatal ultrasound, a task hampered by operator variability and limited labeled data. It leverages USF-MAE, a self-supervised ultrasound foundation model based on a Vision Transformer encoder, fine-tuned for binary VM classification and evaluated with 5-fold cross-validation plus an independent test set. Compared against VGG-19, ResNet-50, and ViT-B/16 baselines, USF-MAE achieves state-of-the-art metrics (validation AUC ~0.96, test AUC ~0.97; accuracy ~97.3%, F1 ~91.8%; high precision ~94–96% and specificity ~98.8–99.3%), underscoring the benefits of domain-specific pretraining for ultrasound. Explainability via Eigen-CAM shows the model focuses on ventricular regions consistent with VM pathology, supporting clinical trust and potential integration into obstetric workflows. The work demonstrates data-efficient, interpretable deep learning for prenatal imaging and sets the stage for multi-institutional validation and extension to 3D or temporal ultrasound data.

Abstract

The proposed study aimed to develop a deep learning model capable of detecting ventriculomegaly on prenatal ultrasound images. Ventriculomegaly is a prenatal condition characterized by dilated cerebral ventricles of the fetal brain and is important to diagnose early, as it can be associated with an increased risk for fetal aneuploidies and/or underlying genetic syndromes. An Ultrasound Self-Supervised Foundation Model with Masked Autoencoding (USF-MAE), recently developed by our group, was fine-tuned for a binary classification task to distinguish fetal brain ultrasound images as either normal or showing ventriculomegaly. The USF-MAE incorporates a Vision Transformer encoder pretrained on more than 370,000 ultrasound images from the OpenUS-46 corpus. For this study, the pretrained encoder was adapted and fine-tuned on a curated dataset of fetal brain ultrasound images to optimize its performance for ventriculomegaly detection. Model evaluation was conducted using 5-fold cross-validation and an independent test cohort, and performance was quantified using accuracy, precision, recall, specificity, F1-score, and area under the receiver operating characteristic curve (AUC). The proposed USF-MAE model reached an F1-score of 91.76% on the 5-fold cross-validation and 91.78% on the independent test set, with much higher scores than those obtained by the baseline models by 19.37% and 16.15% compared to VGG-19, 2.31% and 2.56% compared to ResNet-50, and 5.03% and 11.93% compared to ViT-B/16, respectively. The model also showed a high mean test precision of 94.47% and an accuracy of 97.24%. The Eigen-CAM (Eigen Class Activation Map) heatmaps showed that the model was focusing on the ventricle area for the diagnosis of ventriculomegaly, which has explainability and clinical plausibility.

Figures (7)

• Figure 1: Prenatal ultrasound comparison of normal and ventriculomegaly (VM) fetal brains. (A) Normal fetal brain showing typical lateral ventricle size and configuration b35. (B) Fetal brain with VM demonstrating enlarged lateral ventricles.
• Figure 2: (A) Representative fetal ultrasound image used for analysis b35. (B) Corresponding pixel distribution mapped in HSV colour space, highlighting valid grayscale pixels versus those outside the grayscale range.
• Figure 3: (A) Original fetal ultrasound image. (B) Binary mask identifying non-grayscale artifact regions. (C) Processed image after artifact removal and inpainting using a Navier-Stokes-based b14 inpainting method.
• Figure 4: Overview of the proposed VM classification pipeline. (1) Previously published USF-MAE pretraining framework: A unified ultrasound dataset undergoes preprocessing and is used for self-supervised pretraining with a Masked Autoencoder (MAE), where random patches are masked (25% masking ratio) and reconstructed to learn ultrasound-specific latent representations b13. (2) This work: The pretrained encoder is transferred to the downstream VM classification task by discarding the MAE decoder and attaching a classification head. The model is then fine-tuned on the labeled VM dataset to enable clinical prediction.
• Figure 5: ROC (A, C) and Precision–Recall (B, D) curves for the 5-fold ensemble model on the validation (A, B) and test (C, D) sets. Curves for individual folds are shown along with the mean ensemble performance and variability across folds.