Fetal ultrasound screening learns from scarce labels: RadarMAE's radar-style masking holds cross-hospital metric drops to 0.04
Synopsis
First-trimester ultrasound images for Down syndrome screening are low-resolution and structurally variable, and confirmed rare cases are scarce, so standard self-supervised pretraining struggles. RadarMAE instead masks alternating concentric rings and angular sectors, a radar-like pattern that forces the network to learn the continuous curvature and elongated shapes of craniofacial structures, reaching AUC 0.99 and 97% accuracy on an internal cohort. With domain-adversarial adaptation, the model calibrates on unlabeled images from four external hospitals and limits the drop in key metrics to 0.04 or less.
Interpretation
First-trimester ultrasound screening for trisomy 21 can now be handled by a self-supervised model pretrained on limited clinical data, and it keeps near-internal discrimination across several external hospitals. Manual first-trimester screening reaches only 69% detection at about 92% specificity, well below the roughly 90% detection and 95% specificity of the second trimester; standard self-supervised methods rely on the spatial redundancy of millions of images, while trisomy 21 incidence is below 0.1%, so primary hospitals often see fewer than 10 positive cases a year. On the internal cohort, ViT plus RadarMAE reached AUC 0.99, accuracy 0.97, sensitivity 0.93 and specificity 0.98, above the ResNet transfer-learning baseline that also used spatial registration (AUC 0.98, accuracy 0.95, sensitivity 0.90, specificity 0.96).
The model first learns fetal craniofacial structure from unlabeled images, then is fine-tuned for classification with few labeled samples, sidestepping the need for large-scale annotation. Conventional masked autoencoders mask random patches, which on structure-dense ultrasound images can cover a critical region such as nuchal translucency entirely and leave the network without clinically meaningful representations. RadarMAE masks concentric rings at fixed radial intervals and sectors at angular intervals, and the combined pattern resembles a radar sweep while preserving enough neighboring context for reconstruction; after pretraining the decoder is discarded and the encoder is used downstream.
For cross-hospital deployment, the model needs only a small set of unlabeled images from the target hospital and no local diagnostic labels. Different ultrasound devices, acquisition protocols and operator experience cause domain shift, and primary institutions have too few positive cases to train a supervised model independently. The core-and-branch framework pairs a main diagnostic classifier with a domain classifier whose gradient reversal layer reduces the distinguishability of source and target features; with BOGH as source and four hospitals as targets, each target supplying about 50 to 100 unlabeled images, external aggregate results were AUC 0.98, accuracy 0.95, sensitivity 0.93 and specificity 0.94, and all trisomy 21 cases in the XYH and TAHZU external cohorts were detected.
Perspective
The demonstrated scope is two-dimensional midsagittal fetal face images at 11 to 13+6 weeks of gestation with karyotype confirmed by invasive testing, and all training and evaluation data come from five tertiary hospitals in China; the model is trained and evaluated only for trisomy 21.
External evaluation is presented as aggregate AUC, accuracy, sensitivity and specificity across four hospitals, and per-hospital full metrics are not listed item by item in the main text; the authors also note that turning population-level heatmaps into a real-time interface embedded in routine workflow will require further ergonomic and software refinement.