Kalman Smoothing of Rater Disagreement Across Slices

When several radiologists outline the same organ or tumor on a CT or MRI scan, their outlines disagree. How much they disagree, and who is off, is not constant through the volume. Experts tend to agree in the middle of the prostate or kidney and diverge at the top and bottom of an organ, where partial-volume blur makes the edge genuinely ambiguous. Every current model of this disagreement treats each rater's habits as one fixed setting per rater, or treats each slice independently. We propose to describe a stack of multi-rater annotations as a signal that evolves from slice to slice, and to estimate it with a Kalman smoother, the classical minimum-variance recursive estimator from tracking and brain-machine-interface decoding. Each rater gets a scalar "draws too large / too small" offset and a noise level, and both are allowed to drift smoothly along the slice axis. The smoother returns three things in closed form at negligible cost: a consensus boundary, a calibrated uncertainty band around it, and a per-slice reliability profile for every rater.

The question this answers is empirical and has not been asked: is the structure of inter-observer disagreement in volumetric scans static per rater, or does it vary systematically along the volume? If it varies, a cheap single-decoder network trained on the smoothed targets should match heavy multi-decoder Bayesian models on uncertainty benchmarks at roughly one-seventeenth of their inference latency. If it does not vary, the static model is sufficient, and the field learns that the extra machinery in rater-specific networks buys little.

Core Claim

Represent each rater's mask on slice $k$ by its signed distance map, an image whose value at each pixel is the distance to the drawn boundary, positive outside and negative inside. In that representation, a rater who habitually draws a slightly larger outline adds a roughly constant shift near the boundary. That is exactly the erosion/dilation model used to simulate raters in OMBA-UNet's synthetic benchmark. We write

$$d_k = d_{k-1} + w_k,\qquad b_{r,k} = b_{r,k-1} + u_{r,k},$$

$$z_{r,k} = d_k + b_{r,k}\,\mathbf{1} + v_{r,k},\qquad v_{r,k}\sim\mathcal{N}(0,\sigma_{r,k}^2 I).$$

Here $d_k$ is the latent consensus distance map on slice $k$. $w_k$ is spatially correlated Gaussian change between slices. $b_{r,k}$ is rater $r$'s scalar size offset, drifting by small Gaussian steps $u_{r,k}$. $z_{r,k}$ is that rater's observed distance map, and $\sigma_{r,k}^2$ is the rater's slice-specific noise level, given a log-random-walk prior. The offsets are constrained to sum to zero across raters on each slice, so the consensus is identifiable as the bias-corrected average rater rather than an unknowable truth.

With the noise levels fixed, a forward-backward (Rauch–Tung–Striebel) Kalman pass gives the exact posterior mean and covariance of $d_k$ and $b_{r,k}$. The noise levels and process variances are then updated by expectation-maximization with a variational step, and the two steps alternate. The posterior covariance is the quantity Wu et al.'s cursor-decoding Kalman filter used to express confidence in a moving state. Here it becomes a per-pixel boundary uncertainty that accounts for both rater scatter and slice-to-slice continuity.

Relation to Prior Work

Wu et al. supply the method. Their state-space decoder beat a fixed linear filter on 2D cursor reconstruction (mean squared error 6.28 vs 8.30 cm², trajectory correlation 0.815/0.929 vs 0.756/0.915) because it uses continuity across time steps. We borrow the estimator and its covariance, and replace time with slice index and neurons with raters. Their noted weaknesses do not bite in our setting, because distance maps are continuous rather than spike counts. The closest uses of Kalman filtering inside segmentation, KLDD and KaLDeX, embed Kalman-style updates in network features for single-annotator retinal vessels. Neither has any notion of raters, and neither does label fusion.

The multi-rater segmentation literature motivates the proposal and also defines the competition. Pionono gives each rater a Gaussian in an 8D latent space, with the mean as systematic bias and the variance as inconsistency. That is the static version of our model, learned end-to-end inside a network. OMBA-UNet and the multi-decoder U-Net give each rater its own decoder, at a cost of 0.34 s versus 0.02 s per slice for OMBA-UNet. MRNet uses a recurrent ConvLSTM to fuse expertise cues, but as a feature mixer with no explicit noise model. The variational inference model samples rater-agnostic latents. All of these treat rater behavior as fixed across a volume and work slice by slice or in costly 3D.

Our claim differs in a specific way. Rater parameters become slice-indexed latent states estimated analytically on the label side, before any network is trained.

QUBIQ constrains the design. It reports that uncertainty quantification degrades in 3D tasks under memory limits, and SegTHOR shows that splitting volumes into 20–40-slice packages creates boundary discontinuities. Both push us toward 2D networks plus a cheap continuity model rather than full 3D convolution. The esophageal CTV work supplies evidence that signed distance maps are a workable boundary representation. DiFF argues that consensus is the wrong target and that fusion should serve diagnosis. We do not resolve that tension. Our estimator needs no diagnosis labels, and DiFF's objective could later reweight our consensus.

Operator fatigue in whole-body annotation, noted by qPSMA, motivates drift over the course of annotation. We are explicit, however, that slice order is a spatial ordering, not the order in which a rater worked. Any fatigue claim is therefore tested only in simulation.

Research Program

The first deliverable is the estimator itself. It covers multi-component masks and slices where a rater draws nothing, which we treat as a missing observation that the filter simply skips. It also implements the variational noise-level update. Spatial correlation in $w_k$ is handled by running the filter in a low-dimensional basis (a truncated cosine basis of the distance map near the boundary band), so a 512×512×200 volume smooths in seconds on a CPU.

The second deliverable is a scientific measurement: for each dataset, how much of the disagreement is explained by slice-varying rater offsets and noise rather than static ones. We read this off the learned process variances and the held-out likelihood comparisons below.

The third deliverable tests whether the smoothed posterior is a better training target. A single standard 2D U-Net with two output heads is trained to predict the posterior mean distance map and the log posterior variance under a Gaussian likelihood. At test time, binarizing the predicted distribution at the QUBIQ thresholds gives a soft segmentation at plain U-Net cost.

Baselines

The naive reference point is the per-pixel average of rater masks, which is the soft label most QUBIQ entries train on, together with its majority-vote binarization. It costs nothing and sets the floor for every comparison.

STAPLE is the classical expectation-maximization estimator of a consensus mask and per-rater sensitivity and specificity. It is the closest classical competitor because it already does statistical rater-performance estimation without any network. We run it per volume with the same raters, and beating it means higher held-out-rater likelihood and better calibration of the consensus band.

The ablation that isolates our key ingredient is the same model with $u_{r,k}$ and the noise-level random walk switched off. That leaves a static inverse-variance-weighted fusion of distance maps with one offset and one noise level per rater per volume. This is optimal linear estimation without dynamics, and the label-side analogue of Pionono. Everything else is held identical: estimator code, distance-map basis, and EM iterations.

A second ablation removes the slice-transition term on $d_k$, fusing each slice independently, to measure the separate value of continuity in the consensus.

For the learned comparison, the strongest competitors are OMBA-UNet and Pionono. The file reports no absolute QUBIQ numbers for OMBA-UNet beyond "better than PHiSeg and Probabilistic U-Net on 5 of 7 tasks" and its 0.34 s latency. We therefore anchor on the multi-decoder U-Net, second in QUBIQ 2020, which reports task scores of 0.852 and 0.871 on brain tumor tasks 1 and 2, 0.870 on kidney, and 0.553 on prostate task 1. We also anchor on the QUBIQ summary's 0.80–0.85 average Dice for top teams. On LIDC-IDRI we found no reported absolute result on Lacuna for any method, so we rerun Pionono and OMBA-UNet ourselves.

All learned models use the same backbone (ResNet-34 U-Net, as in Pionono), the same splits, the same augmentation, and an equal hyperparameter search budget of a fixed number of trials each. Beating the strongest method means matching its Q-Dice within its seed-to-seed standard deviation, or exceeding it, at a measured latency near 0.02 s per slice. Lower generalized energy distance would be a bonus.

Why not just…?

Why not just run STAPLE? STAPLE estimates rater performance once per volume. Its parameters are pixel-classification rates, not boundary offsets, so it cannot say "rater 2 overdraws at the organ apex." Pionono's reported κ of 0.758 on Gleason 2019 already beat STAPLE-style fusion in 2D pathology. Whether our model beats STAPLE on volumetric data is not established, and the held-out-slice likelihood test below settles it directly.

Why not just estimate rater parameters in a sliding window of slices? This is the real threat. A windowed static estimator with a well-tuned window could capture most of any slice-varying structure. The Kalman formulation only earns its place if learning the smoothing amount by EM beats the best window chosen by cross-validation, or if its covariance is better calibrated. We include the windowed estimator as a baseline with its window tuned on validation data. If it ties, we will report that the contribution is the measurement of slice-varying rater behavior, not the estimator.

Why not just train Pionono or OMBA-UNet on 3D patches? QUBIQ reports 3D uncertainty quantification degrading under memory constraints. SegTHOR shows discontinuities at package boundaries, and OMBA-UNet is already 17 times slower than a U-Net in 2D. Experiment 3 measures whether a cheap network trained on smoothed targets closes the accuracy gap. If it does not, the heavy models keep their place.

Why not just MC dropout? On QUBIQ brain growth, Bayes dropout scored 0.5083 continuous Dice versus 0.5372 for a structured latent model. Its uncertainty is about the model, not about the raters. We include it only as a calibration reference.

Proposed Experiments

The first experiment uses synthetic raters with known truth. Starting from clean single-annotator volumes in SegTHOR (heart, esophagus) and the QUBIQ kidney task, we simulate 3–7 raters by applying slice-varying erosion and dilation plus boundary jitter, following OMBA-UNet's synthetic protocol but with offsets that ramp toward organ ends. We also include a condition where noise grows with slice index, as a fatigue analogue. We measure three things. The first is consensus boundary error in average surface distance against the clean mask. The second is correlation between recovered and true per-slice offsets. The third is coverage of the 90% posterior band. The dynamic model should beat the static ablation when drift exists and tie it when we simulate static raters. Failing that control would mean the model is overfitting.

The second experiment is the decisive real-data check. On LIDC-IDRI (4 readers, where each reader annotates every slice of a nodule), the QUBIQ volumetric tasks, and the rater-aligned liver set, we hide 30% of the slices of one rater at a time. We then predict that rater's distance maps on the hidden slices from the others' annotations and the rater's own visible slices. The comparison metrics are Gaussian predictive log-likelihood and Dice of the predicted rater mask, across five estimators: dynamic, static, windowed, STAPLE, and mean.

A gain of the dynamic over the static model that is consistent across datasets, with paired bootstrap intervals excluding zero, is the evidence that rater behavior varies along the volume. A null result is an equally reportable finding: per-volume static rater models suffice.

The third experiment tests downstream training. A plain 2D U-Net is trained on dynamic-posterior targets, static-posterior targets, and mean soft labels, and compared against Pionono, OMBA-UNet, and the multi-decoder U-Net on QUBIQ Q-Dice, generalized energy distance, and expected calibration error, with latency measured on identical hardware. The ordering we expect, and would need to see, has dynamic targets above static targets above mean labels. On the tasks with real slice structure, we additionally need the dynamic-target U-Net to fall within the strongest multi-rater network's seed variance. If the target ordering holds but the gap to OMBA-UNet stays large, the value lies in the label model rather than as a replacement for rater-specific networks, and we will say so.