AI model VIFBA transforms fetal brain scans with ultrasound

A team of researchers from Shenzhen University and Harvard Medical School has developed **VIFBA**, a cross-modal ultrasound-video framework that predicts fetal brain ventricular volumes and abnormalities with MRI-like accuracy using only routine ultrasound scans. Published on arXiv, this breakthrough addresses a critical gap in prenatal care where ventriculomegaly (VM) screening currently relies on manual measurements that are operator-dependent and may miss subtle abnormalities.

The system leverages three key innovations: a JEPA-inspired spatio-temporal prediction objective that learns coherent representations from ultrasound videos, a contrastive cross-modal alignment strategy that transfers MRI-derived structural knowledge to ultrasound during training, and a training-free vision-language model for verifying uncertain predictions. When validated on 857 paired cases (3,196 videos), VIFBA achieved remarkable performance metrics including 0.59 mL mean absolute error and 0.99 Pearson correlation for volume regression, 0.94 accuracy for VM severity classification, and 0.78 F1-score for multi-abnormality detection. These results substantially outperform single-task baselines, video-based competitors, and state-of-the-art foundation models.