This post explains our published research: ELLIPSE-YOLO: Automated Fetal Head Circumference Measurement Using Deep Learning Segmentation and Ellipse Fitting (Rai, 2026, DOI: 10.5281/zenodo.19189069).
Every routine pregnancy scan includes a measurement most parents never think about: the circumference of the baby’s head. Obstetricians use it to date the pregnancy, track growth, and catch problems early — it is one of the workhorse numbers of antenatal care. And in most clinics it is still measured the manual way: a sonographer looks at a grainy ultrasound image, decides where the skull outline is, and traces an ellipse over it by hand. Done well, it is accurate. Done at the end of a twelve-hour shift, or by a less experienced operator, or in a rural clinic with no specialist at all, it varies.
Our system automates the whole measurement. The pipeline has two halves that play to different strengths. The first half is deep learning: a YOLO segmentation network — the same family of models that powers modern real-time object detection — is trained to find the skull region in the ultrasound image and outline it as a pixel mask. The second half is deliberately not deep learning: classical geometry fits a mathematically exact ellipse to that mask using constrained least-squares, and the circumference is then computed with Ramanujan’s century-old approximation formula for an ellipse’s perimeter. Neural network for perception, trusted mathematics for measurement — the machine does what machines are good at, and the final number comes from a formula any clinician can verify.
On the international HC18 benchmark, the system detected the fetal head in 100% of test images, with segmentation quality near the ceiling of the metric (mask [email protected] of about 0.995). Just as important for the real world: it runs on ordinary, affordable GPU hardware — no cloud dependency, which matters everywhere connectivity or budgets are thin.
Why does an oncology helpdesk publish obstetric imaging research? Because the underlying discipline is the same one that runs through everything we do: take a medical measurement that depends on tired human eyes, make it reproducible, keep the mathematics inspectable, and publish it openly so others can verify rather than trust. The paper and metrics are free at zenodo.org/records/19189069.