Training Ultrasound Segmenters from the Measurement Lines Sonographers Already Draw

Every routine obstetric ultrasound exam leaves behind a few short line segments: the calipers a sonographer drops to measure a biparietal diameter, a cervical length, or the geometry behind the fetal head's angle of progression during labor. These lines are archived by the million. Pixel-accurate outlines of the same structures, which segmentation models need for training, are almost never drawn. That gap matters most where specialists are scarce, because training a model for a new clinic, device, or anatomy would otherwise require an expert to hand-trace hundreds of noisy images. We propose to train segmentation models for low-cost ultrasound using only these routine caliper lines as supervision.

The method has two parts. A large vision model pretrained on natural photographs, used frozen, proposes candidate outlines from each caliper. We then treat the caliper as a measurement contract: an accepted outline must reproduce the measured length along the measured axis, with the caliper endpoints sitting on its boundary. Outlines that break the contract are rejected or repaired, and the surviving pseudo-masks are distilled into a small network fast enough for bedside hardware. Put plainly, the line a clinician drew to measure something is used twice: once to tell the model where to look, and again to check whether its answer is the right size.

Core Claim

Our claim is that a caliper carries more information than a click. It gives a location and also a metric constraint that any correct mask must satisfy. Combining that constraint with cross-image agreement in frozen foundation-model features should yield pseudo-masks good enough to train a lightweight segmenter. We aim to come within a few Dice points of semi-supervised methods that use dozens of hand-traced masks, on ultrasound structures that rigid shape priors cannot handle.

The contrast with existing work is specific. Weakly Supervised Lesion Co-Segmentation on Ct Scans showed that one-dimensional RECIST lines (the standard tumor diameter markers in CT) can seed full lesion masks, reaching 90.3% Dice on DeepLesion. It did this on high-contrast CT, using GrabCut and a trained co-segmentation network, and used the line only as a seed. GazeRefine steers frozen DINOv3 features with eye-tracking heatmaps to reach 88.10% Dice on polyps without training masks. However, it needs eye-tracking hardware and degrades to 80.28% on low-contrast prostate MRI, which suggests frozen natural features alone will struggle with speckle. ERSR gets 92.05% Dice on HC18 with 10% labels by fitting ellipses to pseudo-labels, but the ellipse is specific to the skull and the 10% are full masks.

Our measurement-consistency check is meant to replace the ellipse with a constraint that applies to any measured structure, including non-elliptical ones like the cervix and pubic symphysis. The frozen features replace eye-tracking with signals already in the clinical record. We do not claim the line-to-mask principle itself is new; the contribution is making it work on ultrasound, where CT's contrast is absent.

Research Program

The pipeline has three stages, and each can be removed for ablation.

The first stage generates candidates. For each image with a caliper (endpoints $p, q$), we derive several prompts: the two endpoints as positive points, the midpoint, and a box whose width along the caliper is $\|p-q\|$ and whose height is swept over a small range. Each prompt goes to SAM (a promptable segmentation model trained on over a billion natural-image masks) and to MedSAM. Each prompt yields a candidate mask, and each image gets about a dozen candidates. This stage alone is the obvious baseline.

The second stage selects candidates by checking them against the caliper and against each other. For candidate mask $M$, let $\ell(M)$ be the length of the chord where the line through $p$ and $q$ intersects $M$, and let $b(M)$ be the mean distance from $p$ and $q$ to the nearest boundary of $M$. The contract score is

$$ s(M) = \frac{|\ell(M) - \|p-q\||}{\|p-q\|} + \lambda \frac{b(M)}{\|p-q\|}, $$

where $\lambda$ weights the endpoint term. Candidates with $s$ above a threshold are discarded. Among the survivors, we keep the one whose mean frozen DINOv2 patch embedding is closest to a class prototype. The prototype is pooled over the highest-scoring masks across the whole training set.

The cross-image step addresses a specific failure. Acoustic shadows produce dark regions bounded by bright edges. A per-image prompt can latch onto such a region while still matching the caliper length by coincidence. The prototype check catches this because shadow interiors do not resemble true anatomy in feature space. Weakly Supervised Estimation of Shadow Confidence Maps in Fetal Ultrasound Imaging documents how such shadows corrupt fetal biometry, which is the motivation for this check.

The third stage distills the selected pseudo-masks into a compact U-Net. Pixels inside candidates that passed the check but disagree with one another are left out of the loss, a partial loss in the style of Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation. The deployed model therefore carries no foundation backbone. This follows the frozen-encoder-plus-light-head logic of FM-OSD and the feed-forward speed of the shadow-confidence network, which ran at 0.0353 s per frame.

Public benchmarks do not contain archived calipers, so we synthesize them from ground-truth masks using clinical definitions:

  • On HC18, the biparietal diameter is the minor axis of the head outline.
  • On FUGC transvaginal images, the cervical length is approximated as a line along the canal between the labeled cervical lips.
  • On PSFHS, the angle-of-progression construction uses the pubic symphysis long axis and its tangent to the fetal head.

Endpoints are jittered with noise. The noise level is calibrated to inter-observer variability, which is substantial in ultrasound: the shadow-confidence study reports human ICC of 0.51 for shadow annotation. These calipers are cleaner than real archived ones, so we treat the synthetic results as an upper bound. If a clinical partner can supply real caliper exports, we will evaluate on a small held-out set of genuinely archived measurements paired with expert masks traced only for evaluation.

Baselines

The naive reference point is the caliper itself, turned into a mask by fixed geometry: an ellipse on the caliper as one axis with a dataset-average aspect ratio. This tests whether any learning is needed. It should be competitive on the fetal head and poor on the cervix and pubic symphysis.

The strongest training-free competitor is SAM or MedSAM prompted directly with caliper-derived points and boxes, keeping the single highest-confidence output. Fine-tuned SAM held 0.857 Dice on noisy handheld POCUS while a domain-specific model fell to 0.667, as reported by Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound. That study, though, fine-tuned on 7,460 annotated studies and used manual box prompts. Lacuna has no reported result for SAM prompted with caliper-derived prompts on HC18, FUGC, or PSFHS. We will run it with exactly the prompt set our first stage uses. Our method must beat it in Dice and 95th-percentile Hausdorff distance (HD95) on at least the two non-elliptical tasks, or the selection stage has not earned its place.

The strongest label-hungry competitors define the target we must approach:

  • On FUGC, top semi-supervised entries with 50 hand-traced masks reached 90.26% mean Dice, per FUGC.
  • On HC18, ERSR reports 92.05% Dice at 10% labels, against a plain U-Net's 70.89%.
  • On PSFHS test set 2, fully supervised challenge entries reached fetal-head Dice of 0.9025 to 0.9327 and pubic-symphysis Dice of 0.7938 to 0.8717, per the PSFHS Challenge Report.

We will retrain ERSR and a Mean-Teacher model on matched data rather than only quote these figures.

Two further comparisons isolate the key ingredient:

  • An nnU-Net partial-loss model that treats the caliper line pixels as foreground scribbles and a band far outside the caliper as background. This is the Gotkowski recipe, which reached 0.813 mean Dice on ScribbleBench, a CT/MRI benchmark rather than ultrasound.
  • Our full pipeline with the contract score $s(M)$ removed, keeping only prototype selection, and separately with prototype selection removed.

All learned methods share the same student architecture, input resolution, augmentation, and hyperparameter search budget. Beating a competitor means a higher Dice with a 95% bootstrap interval that excludes zero difference, and no worse clinical measurement error.

Why not just…?

Why not just prompt SAM with the caliper endpoints and stop there? This is the real threat, and we expect it to match our method on the fetal head, where the bright skull ring gives SAM a clear edge. The open question is the cervix and pubic symphysis. There, a point or box prompt has several plausible outlines, and the single-image confidence SAM reports does not encode size. The second ablation settles it. If prompted SAM is within one Dice point of the full pipeline on all three tasks, the contribution reduces to distillation for speed, and we will report it that way.

Why not just train nnU-Net with partial loss on the caliper lines as scribbles? The Gotkowski result is strong, but ScribbleBench scribbles cover the object's interior and span its extent. A caliper is one chord touching the boundary only at two points, so the network receives almost no information about the boundary away from those points. We expect this baseline to leak along shadow edges. Because it is run head-to-head under identical student and budget, any advantage it has will show up directly.

Why not just pay a sonographer to trace 50 masks? For a single well-resourced project, that is cheaper, and FUGC shows it works. The argument for calipers is about scale and transfer: every new device, site, or anatomy would need its own 50 tracings, while calipers already exist. Hand-traced masks are also noisy in ultrasound, as the shadow-confidence study's human Dice of 0.5443 against reference shows. We will report how many traced masks the caliper-trained model is worth by sweeping FUGC's labeled count from 5 to 50 and finding the crossover point.

Why not just use an ellipse prior as ERSR does? ERSR's authors note it does not transfer beyond the skull. Our naive geometric baseline measures exactly how much the ellipse gives on each task, so the tasks where it fails are documented rather than assumed.

Proposed Experiments

The main experiment trains on caliper-only versions of HC18, FUGC, and PSFHS and evaluates against the held-out expert masks. We report three kinds of metrics:

  • Segmentation quality: Dice and HD95.
  • Clinically relevant measurements: head circumference error in millimetres, compared with the 1.81 ± 1.65 mm reported by Confident Head Circumference Measurement under full supervision, and angle-of-progression error in degrees on PSFHS, where challenge teams averaged under 11°.
  • Deployment cost: student latency on a CPU-class device, against FUGC's 200 ms real-time threshold.

Three further checks decide whether the claim holds.

The first is a robustness sweep over caliper noise and over the fraction of images with calipers. It shows whether the contract score degrades gracefully or collapses when endpoints are off by a few millimetres, which is typical of real archived calipers.

The second is a shadow-stratified evaluation. We split test images by severity of shadowing, using the confidence-map approach of the shadow-confidence study as an automatic grader. Prototype selection should reduce boundary leakage specifically in the heavily shadowed stratum. If the gain is uniform across strata, the shadow explanation is wrong.

The third is a cross-device transfer test. We train on PSFHS images from one ultrasound machine and test on the other. This probes whether frozen natural-image features carry the robustness seen in the cardiac POCUS comparison when no dense masks are used. It is the regime that matters for clinics using equipment unlike that of the hospitals the data came from.