Sonographer Disagreement at the Placental Base as an Early Marker of Preeclampsia

When several trained sonographers outline the same first-trimester placenta on ultrasound, they disagree. Every placenta segmentation method in the literature treats that disagreement as noise. Methods average it into a consensus mask, beat it with a network, or cut the ambiguous pixels away. This proposal tests the opposite hypothesis. Disagreement concentrated at the placental base, where the placenta meets the uterine wall, may partly reflect the tissue itself. That base, called the utero-placental interface, is where preeclampsia is thought to begin. Fetal cells there fail to invade deeply enough to remodel the mother's spiral arteries into wide, low-pressure vessels. The narrow arteries that remain fire blood into the placenta at high velocity and damage the villous tissue nearby. An interface disrupted in this way should look less distinct on ultrasound, so experts should disagree about where it lies.

The core claim is narrow and testable. We separate rater disagreement into three parts: each rater's personal contouring habit, ambiguity explained by known acoustic physics (depth, posterior position, maternal body mass), and a remaining image-intrinsic part. The remaining part, measured at the basal interface in 11–14-week scans, should predict later preeclampsia beyond what standard first-trimester screening captures. Standard screening means placental volume, uterine artery Doppler, and maternal factors. If the claim holds, it gives an in-vivo spatial readout of the early implantation defect. That matters because nearly all mechanistic findings in preeclampsia come from placentas delivered after the disease has already appeared.

Why this gap, and why not the obvious idea

The preeclampsia literature supplies the motivation. Mechanistic studies locate the failure in trophoblast invasion and spiral artery remodeling. Reduced integrin ITGB3 impairs trophoblast migration and adhesion (ITGB3 is reduced in pregnancies with preeclampsia…). SerpinA5 suppresses invasion (Plasma SerpinA5 in conjunction with uterine artery pulsatility index…). The A2M–LRP1 axis is tied to failed artery remodeling (Targeting A2M-LRP1…). Necroptosis via BIRC3 appears in delivered placentas (Integrated multi-omics analysis reveals necroptosis-related biomarker BIRC3…). Direct intervillous flow measurement links high-velocity inflow jets to small-for-gestational-age babies (Direct measurement of blood flow into the intervillous space…).

These studies share two limits. They mostly profile third-trimester tissue, after disease onset. Their clinical imaging counterpart, uterine artery pulsatility index, measures resistance upstream of the uterus and has no spatial resolution at the interface.

The segmentation literature supplies the method and also defines the contrast. Andreasen et al. benchmarked a Mask R-CNN against 25 clinicians and showed systematic tier differences in agreement with the reference: experts 0.69 IoU, residents 0.66, midwives 0.59, model 0.75. They used a consensus reference and did not relate disagreement to outcome. The Medical Image Analysis placenta study compared Monte Carlo dropout uncertainty maps with human inter-rater variability, but only in healthy volunteers. It also showed that acoustic geometry dominates: a U-Net trained on anterior placentas fell to 0.26 Dice on posterior ones. PRISM reaches 0.9715 Dice on 124 volumes at 11–14 weeks and cites preeclampsia as motivation, but never tests outcome. Hempstead et al. prune low-confidence boundary pixels to stabilize tissue-texture measurements. That deliberately discards the zone we propose to study. UPI-Net detects the interface contour, but targets placenta accreta, the opposite defect (too-deep invasion), against a single reference.

Two multi-rater methods supply the machinery. Pionono gives each annotator a learned Gaussian in a latent space. Its mean captures that rater's systematic bias and its variance their self-inconsistency. It was tested only on histopathology, where lighting is uniform. DiFF learns per-rater weight maps that make the fused mask best for a downstream classifier, raising diagnostic AUC by up to 2.73% on REFUGE-2.

The obvious intersection idea is to apply DiFF to placentas so that the fused mask best predicts preeclampsia. We do not make that our core. DiFF still produces a single mask and treats disagreement as something to resolve. It ties the mask to one classifier, a coupling its authors flag as a limitation. With about a hundred preeclampsia cases, an end-to-end diagnosis-guided segmenter would overfit. It would also yield no interpretable statement about where the placenta is abnormal. Our proposal differs in kind. Disagreement becomes the measured quantity, decomposed with explicit physical covariates, localized anatomically, and tested as a pre-specified feature. We keep DiFF as a baseline.

Research Program

The first component is data. We will run a nested case-control study inside a first-trimester screening cohort. Women already undergo 11–14-week combined screening, and we add a standardized 3D placental sweep. Preeclampsia cases (target 120, split into early and late onset) are matched about 1:3 to normotensive controls by gestational age, scanner, and site. Suitable host cohorts are a screening cohort of the Shenzhen Longhua type, which recruited 5,000 women and observed 60 cases, or a site within PRECISE, which follows 6,932 pregnancies with ultrasound and biobanking. Either would need roughly 3,000–6,000 enrolled women to reach the case target.

Five raters, blinded to outcome, outline each volume on a fixed set of slices: two fetal medicine experts, two residents, and one sonographer. They mark the outer placental boundary and, separately, the basal interface. A 20% subset is re-annotated after four weeks to measure intra-rater consistency. Annotators also record a confidence score per slice.

The second component is a model of disagreement. We extend Pionono to 3D ultrasound with three additions. First, each rater's latent distribution captures personal style, as in the original. Second, an explicit acoustic branch conditions on depth from the probe, anterior or posterior position, maternal body mass index, and an image-derived attenuation estimate. It is trained to explain the disagreement that these physical factors predict. Third, an image-intrinsic ambiguity head predicts the remaining per-voxel disagreement directly from the image, which makes the measure usable at deployment with a single scan and no raters.

All three parts are fit only on controls and on cases from training folds. The phenotype is the mean residual ambiguity in a 3–5 mm band around the predicted basal interface, normalized by the same quantity at the chorionic (fetal-side) surface. The normalization cancels global image quality.

The third component is the outcome test. The pre-specified primary analysis adds this interface-ambiguity score to a reference model of maternal factors, uterine artery pulsatility index, and consensus placental volume. It reports the change in AUC and net reclassification for preterm preeclampsia under subject-level nested cross-validation, with matching sets kept intact.

A secondary analysis tests the mechanistic link. A subset of participants gets a late-pregnancy scan within two weeks of delivery. For them, we relate interface ambiguity to placental histopathology scored for maternal vascular malperfusion, and to delivery-tissue ITGB3 and BIRC3 immunostaining. This tests whether the image signal tracks the biology the mechanistic papers describe.

Baselines

The naive reference point is the clinical model alone: maternal risk factors plus uterine artery pulsatility index, refit in our cohort. The closest published number is from Zhang et al.. SerpinA5 combined with pulsatility index and body mass index reached AUC 0.946, and SerpinA5 alone reached 0.881. That result came from a 48-versus-48 development set and is likely optimistic. Where plasma is banked, we will measure SerpinA5 and include it, so the reference model is as strong as the literature allows. We will report our own refit AUC rather than compare to 0.946 directly.

The strongest imaging baseline is consensus placental volume, the biomarker PRISM targets. It is computed from a PRISM-style interactive segmentation run to its 0.95-Dice stopping criterion, plus an automated 3D U-Net. Both models are trained on the same folds, with the same augmentation and the same hyperparameter search budget as our model.

The strongest competing uncertainty baseline is single-model Monte Carlo dropout uncertainty in the same interface band, following the Medical Image Analysis study. It is trained on consensus masks only, with matched compute.

The texture baseline extracts radiomic and quantitative-ultrasound features from the same interface band. Raw radiofrequency data are used where available, following Hempstead et al., and the features feed an elastic-net classifier tuned with the same cross-validation.

The diagnosis-first baseline is DiFF trained to predict preeclampsia from the five-rater annotations. On placentas there is no reported result found on Lacuna for DiFF. Its glaucoma, thyroid, and melanoma AUC gains of up to 2.73%, 2.51%, and 2.04% indicate the size of gain multi-rater calibration has produced elsewhere.

Ablations isolate the key ingredients. The first uses raw pixelwise rater disagreement with no style or acoustic adjustment. The second removes only the acoustic branch. The third moves the measurement band from the basal interface to the chorionic surface, which serves as a negative control.

The proposal counts as beating the baselines under three conditions. The interface score must improve AUC over the reference model plus volume, with a 95% confidence interval excluding zero. It must remain significant when the texture and dropout features are already in the model. And the chorionic-band control must show no comparable signal.

Why not just…?

Why not just use model uncertainty from one network instead of paying five raters? This is the most serious threat, and it is cheap. If Monte Carlo dropout uncertainty at the interface predicts preeclampsia as well as residual human disagreement, the multi-rater apparatus is unnecessary for deployment. The prior placenta study found the two are related, though it did not show they are interchangeable. Our design uses raters only for training, so the deployed score already needs one scan. The head-to-head comparison inside identical folds settles the question. If dropout matches, we will report dropout as the practical tool and keep human disagreement only as its validation.

Why not just compute texture features in the basal region? If the ambiguity is truly in the image, texture should capture some of it. Whether texture captures all of it is an open question. Hempstead et al. show that texture measures become unstable precisely near ambiguous boundaries, and depth confounds texture as strongly as it confounds human judgment. We test this by fitting the interface score with texture already in the model. If the score adds nothing, the honest conclusion is that disagreement mapping is a way to localize where to measure texture, not a new signal. We would then reframe the contribution in those terms.

Why not just fuse raters with STAPLE or majority vote and use the resulting volume? Volume from consensus masks is a baseline, not a strawman, because volume is the biomarker the field already pursues. Andreasen et al.'s tier differences show that an unweighted vote mixes in rater habits, and Pionono showed that modeling raters beats STAPLE on histopathology. Neither result says anything about outcome. The comparison of volume against volume plus interface score answers that directly.

Why not just apply DiFF so the mask is optimized for preeclampsia? DiFF is included as a baseline. With about a hundred cases, a segmentation network trained through a diagnosis loss risks fitting the outcome through the mask. It also yields a mask whose meaning depends on the classifier. Our score is fixed before outcome labels are used, which makes the test cleaner and the result interpretable as a statement about the interface.

Decisive Checks

Three results would falsify or confirm the idea. The first concerns the acoustic adjustment. In the 12-patient multi-rater subset of the Medical Image Analysis study, and in our controls, the acoustic branch should explain most of the anterior-versus-posterior gap in disagreement. If it does not, the residual is still contaminated by acoustic physics and the outcome analysis is uninterpretable. This check runs before any outcome is unblinded.

The second is the primary test: the pre-specified gain in AUC and net reclassification over clinical factors, pulsatility index, volume, texture, and dropout uncertainty. It should be larger for preterm than for term preeclampsia, because preterm disease is the subtype most tied to failed artery remodeling.

The third is anatomical specificity together with biological coherence. The chorionic-band control must be null. In the late-scan subset, interface ambiguity must correlate with maternal vascular malperfusion scores and reduced ITGB3 staining at the basal plate.

A positive primary result with a null specificity check would mean we have measured image quality, not placental biology. A null primary result with good adjustment would also be informative. It would show that first-trimester expert disagreement is not a window on the implantation defect, and that the field is right to treat it as noise.