Moein Heidari, Junbo Rao, Yudan Chen, Yasamin Medghalchi, River Jiang, Wenjin Chen, David J. Foran, Jasmine Grewal, Ilker Hacihaliloglu
8 min
Abstract
**Purpose:** Ultrasound biometry is widely used across various clinical applications, including prenatal and cardiac examinations, to obtain quantitative measurements by placing calipers at anatomical landmarks. However, these measurements remain subject to interoperator variability as they rely largely on manual caliper placement. Conventional methods often use specialized prediction heads for different ultrasound tasks, whereas vision-language models offer flexibility through text instructions. However, direct coordinate prediction does not explicitly model spatial likelihood distributions, limiting the representation of localization ambiguity. **Methods:** We propose Kev-Bio, a vision-language model that uses clinical measurement questions to guide ultrasound landmark detection. Inspired by System One decision models, Kev-Bio uses decision representations of the language model to generate queries for each caliper endpoint. These queries are matched against shared dense ultrasound features to produce spatial probability maps, from which landmark coordinates and clinical measurements are derived. The model is trained end-to-end using both landmark localization and clinical measurement losses. **Results:** On the nine-view FU/GU Biometry 2026 dataset, Kev-Bio achieves a mean radial error (MRE) of 22.50 pixels and a clinical measurement mean absolute error (MAE) of 15.38, outperforming the compared baselines in overall performance. **Conclusion:** These results demonstrate the effectiveness of Kev-Bio for ultrasound landmark detection and clinical measurement across different anatomical views. Our code and models are available at https://github.com/moeinheidari7829/Kev-Bio.
Sam: Here, System One refers to scoring candidate answers rather than generating an answer sequentially. The authors build on the Kev decision model. They adapt its decision representations to a spatial problem: which location best answers this clinical localization question?
Alex: Walk through an actual question. What changes when the same image supports different measurements?
Sam: Take an apical four-chamber image, a heart view showing the atria and ventricles. Asking for ventricular length requires different endpoints from asking for atrial width. The paper shows those questions producing different probability maps on the same image, concentrated around the corresponding landmarks.
Alex: How does a sentence become something that can search an image? That's the bridge I'd want to understand before trusting the result.
Sam: The model processes the ultrasound image, a description of its anatomical view, and the measurement questions. Each question produces an internal decision representation. A small network turns that representation into a separate query for each required endpoint.
Alex: But the queries need an image representation detailed enough to distinguish nearby boundaries. Where does that spatial detail come from?
Sam: The model combines features from a language-aligned vision encoder, an ultrasound foundation model, and DINOv2, another visual encoder. A fusion module aligns their features into a shared spatial map. It also incorporates shallow image features to preserve fine detail.
Alex: So the question determines what to look for, while the shared image map supplies possible locations. How does the model choose among them?
Sam: Each endpoint query is matched against that map. Stronger matches receive higher probabilities. The predicted coordinate is a probability-weighted average of locations, rather than simply the strongest location. The endpoints then feed a geometric calculation for the requested measurement.
Alex: That average could land between alternatives when a map has separate peaks. Does the paper show that kind of ambiguity?
Sam: It does. An angle-of-progression example shows competing candidate locations, with the annotation between them. That's why retaining the map matters: a coordinate alone would hide the competing possibilities. Training supervises the maps, the endpoint coordinates, and the measurements calculated from those coordinates.
Alex: Let's turn to the evidence. How broad was the evaluation, and what was actually held out?
Sam: They used about sixty-seven hundred labelled images from the FU/GU Biometry benchmark. These span nine ultrasound views across cardiac, prenatal, and intrapartum imaging. They used five-fold cross-validation, with every method trained and evaluated on the same partitions.
Alex: Was that because there wasn't a separate test set available, or was cross-validation an additional check?
Sam: The official challenge test annotations weren't available for offline evaluation. So these results come from rotating held-out portions of the labelled training set. The authors also didn't use the additional unlabelled challenge frames, although some competing methods did.
Alex: What's the headline comparison? Give me the result that best captures whether the landmarks actually improved.
Sam: Average landmark error was about twenty-two and a half pixels. The strongest baseline on that metric was about twenty-four pixels. Those are averages across views, not a guarantee about any particular measurement.
Alex: Pixel error alone doesn't tell me whether a clinical measurement is acceptable. Did the authors evaluate the measurements too?
Sam: They did, and Kev-Bio had the lowest average measurement error among the compared methods. But it wasn't best everywhere: competitors had lower measurement errors for femur length, cervical length, and angle of progression. The reported results establish comparative benchmark performance, not a clinical acceptance threshold.
Alex: The architecture has several ingredients. What shows that spatial matching matters, rather than simply having a language model and several image encoders?
Sam: The component removal study is revealing. Replacing the pixel pointer with direct coordinate regression raised landmark error to about thirty-six pixels. Removing the language model caused a smaller deterioration. My reading is that explicit spatial matching is the strongest demonstrated contribution, with language conditioning adding a further benefit.
Alex: And what about the promise of showing doubt? Did they test whether the probability maps actually identify mistakes?
Sam: They ranked predictions by entropy, a measure of how spread out the probabilities are. Reviewing the most uncertain ten percent captured forty percent of localization errors. Here, an error meant a deviation beyond a threshold defined relative to the image diagonal.
Alex: That supports prioritizing review, but it doesn't mean the probabilities are calibrated measures of correctness. Nor does it show that clinicians using them make better decisions.
Sam: Those conclusions aren't established here. The authors identify uncertainty calibration as future work. Their main limitation is broader still: evaluation covers predefined measurements on a single benchmark. Unseen instructions, new anatomical structures, and different imaging conditions remain untested.
Alex: Then who should read the full document, and where should they start?
Sam: Researchers building ultrasound measurement systems or language-guided localization models. Start with “Language Conditioned Pixel Pointer,” then the component ablation comparison. Read “Discussion and Limitations” before treating the shared framework as evidence of generalization. The code and models are available for closer investigation.
Alex: And for someone outside that area, what's the idea to carry away?
Sam: Don't just predict where a measurement belongs; preserve the competing locations too. In this benchmark, that spatial approach improved localization and helped identify predictions worth checking. Whether it transfers beyond these predefined tasks is the next question, not an answered one.
Alex: Thanks for listening.