Researchers have built an AI system that estimates fetal brain ventricle volume from a routine ultrasound video, without needing an MRI scan.
The system, called VIFBA, is trained on 857 paired ultrasound-and-MRI cases (3,196 videos total). During training, it learns to align structural detail from MRI scans with patterns in ultrasound video; at actual use, it only needs the ultrasound. On held-out test data, the model predicted ventricular volume with a mean error of 0.59 mL and correlated with true MRI measurements at 0.99, while classifying ventriculomegaly severity with 94 percent accuracy. A separate module, built on a vision-language model, flags other potential brain abnormalities beyond ventricle enlargement, reaching an F1 score of 0.78.
Ventriculomegaly, the enlargement of brain ventricles, is currently screened by manually measuring width on a 2D ultrasound plane, a method that varies by who's holding the probe. MRI gives a cleaner volumetric read but costs more and isn't available everywhere pregnant patients get care. A tool that gets MRI-level detail out of the ultrasound machines clinics already have could extend better screening to lower-resource settings, not just wealthier hospitals with MRI on-site.
The usual caveat applies: this is a single-institution research result validated on one dataset, not a clinical tool yet, and getting from a strong arXiv paper to something a sonographer trusts in the room takes years, not months.