Paper review · November 5, 2026
Every Pixel Is a Choice: A System One Model for Medical Ultrasound Biometry
Every Pixel Is an Option: A System One Model for Medical Ultrasound Biometry
It encodes clinical measurement questions with a language model to build per-endpoint queries, matches them against fused ultrasound features to produce probability maps that locate caliper positions, and achieved lower error than existing methods across 9 views.
Problem
Ultrasound biometry (fetal head circumference, cardiac dimensions, etc.) is done by placing calipers manually, so results vary from examiner to examiner. Existing automated models use a separate output head for each task, which makes them hard to extend to new measurements. Approaches in which a VLM outputs coordinates directly cannot express uncertainty about location.
Method
Kev-Bio feeds ultrasound image tokens (from the Qwen vision encoder and USFM), a view description and a measurement question into Qwen3.5-0.8B initialized with the Kev adapter. It uses the hidden state of the <decide> token at the end of each question as the decision representation, and builds a query for each caliper endpoint with an MLP and endpoint embeddings. Qwen3.5, USFM and DINOv2 features are fused by a Pixel Neck into a shared high-resolution feature map. A spatial softmax over the dot product of the queries and the feature map yields probability maps, and soft-argmax computes the coordinates. Measurement values are derived with geometric functions for distance, angle and circumference. The model is trained end-to-end with a combination of coordinate L1, Gaussian heatmap cross-entropy, measurement error and auxiliary regression losses.
Results
- On the FU/GU Biometry 2026 data (9 views, 6,743 images), 5-fold cross-validation gave a mean MRE of 22.50px, lower than the strongest baseline, CAMS (23.88px).
- Mean MAE on clinical measurements was 15.38, lower than FMUB (17.47). It was the most accurate in landmark location on 8 of 9 views and in measurement values on 6 views.
- On the fetal head (HC) view, MRE fell from IMPACT-L's 17.03px to 13.76px.
- Using the entropy of the probability map as an uncertainty score, the 10% most uncertain predictions contained 40% of all errors (versus 10% under random selection). Error-detection AUROC was 0.70 for cardiac, 0.76 for prenatal and 0.86 for intrapartum.
Limitations
Because the official test labels were unavailable, evaluation used only 5-fold cross-validation on the training set. The data is highly imbalanced across views (38 images for IVC, 4,000 for AoP). Cardiac views have few samples. For femur length, cervical length and the angle of progression, measurement error was slightly higher than other methods. The roughly 190,000 unlabeled frames provided were not used.
Why it matters
Because a single model handles multiple ultrasound measurements just by changing the text question, adding new measurement tasks is relatively easy. The uncertainty score based on probability maps can be used to flag caliper positions that clinicians should re-check, which makes it meaningful as an aid in actual image reading.
Key terms
- Ultrasound biometry
- An examination that measures quantitative values such as distance, angle and circumference by placing calipers on anatomical landmarks in ultrasound images.
- System One decision model
- A model that does not generate answers autoregressively but instead scores candidate responses and picks one. This paper treats each pixel as a candidate and scores it.
- Soft-argmax
- A differentiable method that computes coordinates as the expected location over a probability map. It makes end-to-end training possible.
- MRE(Mean Radial Error)
- The mean Euclidean distance between predicted and ground-truth landmarks. This paper reports it in pixels.
Original
Authors: Moein Heidari, Junbo Rao, Yudan Chen, Yasamin Medghalchi, River Jiang, Wenjin Chen, David J. Foran, Jasmine Grewal, et al. · Posted 2026-11-02 · License CC BY-NC-SA 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.