Image-Derived Fuzzy Ambiguity as a Label-Free Calibration Target Across Hospitals

Segmentation networks for medical scans tend to fail in a predictable place. The failures sit on low-contrast boundaries, such as a thin heart wall or a pancreas pressed against bowel, and they get worse when the scan comes from a hospital or scanner the model never saw. The network is usually just as confident on these wrong boundary pixels as on the easy interior. Penalties that make the model less certain on its own mistakes do help, but they need ground-truth masks to know where the mistakes are, and masks are exactly what we lack for unlabeled scans and external hospitals. This proposal borrows a label-free signal from classical image processing. Fuzzy entropy maximization is an old thresholding technique: it fits soft "dark versus bright" membership curves to an image's intensity histogram and chooses the curve parameters that make the partition as informative as possible. Pixels whose membership ends up near one half are the ones the image itself cannot assign to either side.

The idea is to compute that image-intrinsic ambiguity locally around each predicted boundary, on every scan, including scans from new hospitals. We then train the network so that its uncertainty tracks this ambiguity: soft where the image is ambiguous, crisp where the image is clear. The signal comes from the pixels, not from the network or from labels, so it is recomputed under each scanner's own contrast. It also cannot drift the way a model's self-estimated uncertainty does during teacher-student training. If this works, it gives a calibration and consistency target for exactly the scans where label-based methods cannot run.

Core Claim

The claim is that a locally fitted fuzzy-entropy partition of image intensities predicts where a segmentation network will be wrong. On out-of-center scans, it should predict this better than the network's own predictive entropy. Matching the network's per-pixel entropy to this target on unlabeled multi-center data should then improve boundary calibration and boundary distance metrics on an unseen vendor, beyond the strongest mixed-domain semi-supervised method.

This also takes a position on a tension in the literature. One line of work sharpens boundaries: Zheng et al. introduce an overlap loss that pushes foreground and background probabilities toward 0 and 1, which raised left-atrium Dice from 82.01% to 92.24%. Another line softens them: MEEP maximizes entropy on misclassified pixels, and Cheng et al. maximize the spectral entropy of logits, and both argue that crispness at boundaries causes overconfidence. Our target says the answer differs per pixel and should be decided by image evidence, and the experiments test whether that resolves the tension.

Research Program

The first step builds the ambiguity map. Take the network's current predicted boundary and place a small window around each boundary location. In that window, fit a two-class fuzzy partition of intensities: an S-shaped membership function whose two or three parameters maximize the total fuzzy entropy of the window histogram, following the maximum-fuzzy-entropy formulation of Tian et al. and Jin et al.. Both papers showed this separates subtle transitional grey levels without supervision. Those works searched the parameters with genetic algorithms and ant colonies. With only a few parameters per window, an exhaustive coarse grid on the GPU suffices, which removes the search-cost problem noted for this line of work.

Each voxel $x$ then receives a membership $\mu(x)$, and its image ambiguity is the binary fuzzy entropy $a(x) = -\mu\log\mu - (1-\mu)\log(1-\mu)$, normalized to $[0,1]$. Two properties of this construction matter. The partition only separates darker from brighter subpopulations, so it does not care which tissue is bright. And it is refit per scan, so it follows each center's local contrast rather than a global intensity style.

The second step uses the map in training. On every unlabeled voxel inside the boundary band $B$, we match the network's normalized predictive entropy $H(p(x))$ to the image ambiguity:

$$ L_{\text{amb}} = \frac{1}{|B|}\sum_{x \in B} \big(H(p(x)) - a(x)\big)^2 . $$

The map $a(x)$ is held fixed, with no gradient flowing through it. Where $a$ is low, this loss sharpens predictions, in the spirit of Zheng et al. Where $a$ is high, it softens them, in the spirit of MEEP, but without needing to know which pixels are errors.

We also use $1-a(x)$ to weight the teacher-student consistency loss on unlabeled scans. As a result, pseudo-labels are trusted less exactly where the image cannot support them. DyCON does something similar using the model's own entropy, so this weighting directly tests our signal against the model's self-estimate.

The whole module is added to an existing mixed-domain pipeline rather than replacing it. It sits on top of MiDSS, which builds intermediate domains between centers through copy-paste and Fourier low-frequency amplitude mixing, and on top of a plain Mean Teacher. Keeping it additive isolates its contribution.

The third step is analysis. We ask whether $a(x)$ predicts real errors, where it fails, and whether the matching loss improves calibration in the boundary band specifically, not just globally.

Baselines

The naive reference is a supervised network trained only on the few labeled source-center scans, evaluated on the unseen vendor. On M&Ms, Campello et al. report that such baselines lose about 6% Dice when moving to an unseen vendor. That paper also supplies the self-configuring reference: nnU-Net pipelines with heavy intensity augmentation won the challenge, reaching about 0.88–0.92 LV Dice across vendors and closing the cross-vendor gap to 1.5%. Those numbers were obtained with full labels, so they serve as an upper reference rather than a direct competitor. We will rerun nnU-Net on our label-scarce split with its default augmentation.

The strongest competitor is MiDSS, because it is designed for exactly our setting: labels from one domain, unlabeled scans from others. It reports a 13.57% Dice improvement over prior methods on prostate and 87.66% Dice on fundus against an 88.46% fully labeled upper bound. No reported per-vendor M&Ms result was found on Lacuna for MiDSS, so we will reproduce it with the released protocol. Our module will be added to that same reproduction with identical data splits, backbone, iterations, and a hyperparameter budget matched trial-for-trial.

DyCON is the strongest entropy-guided semi-supervised method. It uses the model's own voxel entropy to schedule consistency learning and reports 61.48% Dice against 53.53% for the earlier BCP method on ISLES'22 with 5% labels. We run it on the same splits.

For calibration, we compare against two supervised entropy regularizers. MEEP is applied to the labeled portion only. The spectral-entropy method of Cheng et al. reports reducing ECE (expected calibration error, the gap between confidence and accuracy) on BraTS 2020 from 9.1 to 1.5. We also include temperature scaling fitted on source-center validation data.

Beating the field means two things on the unseen vendor and the cross-center prostate split. Boundary-band ECE must be lower than MiDSS's and both supervised regularizers'. HD95 must be lower than MiDSS's. Dice must not fall, and all differences must hold over at least three seeds.

Two ablations isolate the key ingredient. The first replaces $a(x)$ with the network's own predictive entropy, which is essentially DyCON's signal. The second replaces $a(x)$ with a purely geometric target: high entropy within a fixed distance of the predicted boundary, low outside, with the band area matched to the fuzzy map. If the geometric band does as well, then the image intensities carry no extra information, and the claim is wrong.

Why not just…?

Why not just use the network's own entropy, as DyCON does? Its entropy is computed by the same model whose boundaries are miscalibrated. MEEP's authors explicitly flag that under severe domain shift a model may "revert to overconfident wrong predictions." When that happens, self-entropy is lowest exactly where errors are worst. Our first decisive check measures this directly by comparing how well each signal detects errors on the held-out vendor. If self-entropy detects errors as well as $a(x)$, the proposal loses its reason to exist.

Why not just soften a fixed-width band around the predicted boundary? This is the most serious threat. A large share of boundary error may be explained by distance to the contour alone, with no intensity information needed. The area-matched geometric ablation tests exactly this. We commit to reporting it as a negative result if it matches the fuzzy map. If that happens, the honest conclusion is that a cheap geometric prior suffices.

Why not just harmonize intensities or use temperature scaling? Intensity augmentation was the strongest generalization lever in M&Ms. But global harmonization changes brightness and style, whereas our target concerns local edge contrast. CrossPan found that histogram matching gave negligible gains under contrast-mechanism shifts (Peng et al.). Temperature scaling lowers confidence everywhere, including correct interiors, which is the behavior MEEP was designed to avoid. Both are run as baselines, and boundary-band ECE will show whether they already close the gap.

Why not just use a foundation model? On CrossPan, MedSAM2 reaches only about 0.47–0.59 zero-shot Dice. That is far below in-domain performance, and it gives no calibrated uncertainty at boundaries. A foundation model could serve as the backbone for our loss, but on its own it does not answer the calibration question.

Proposed Experiments

The main benchmark is M&Ms. Labels come from 5% and 10% of vendor A scans, unlabeled training scans come from vendors B and C, and vendor D (Canon) is held out entirely. This mirrors the mixed-domain protocol of MiDSS. The multi-center prostate split used by MiDSS is the second benchmark.

The left-atrium MRI dataset, with its standard 80/20 split, serves as a single-center, low-contrast sanity check. It lets us compare directly with the tension between Zheng et al.'s sharpening loss and MEEP's softening loss.

CrossPan's T1W→T2W transfer is a deliberate stress test. When contrast between tissues largely vanishes, $a(x)$ should become high almost everywhere. We predict the method will improve calibration and the flagging of uncertain regions but will not rescue Dice from collapse, and we will report the outcome either way.

Metrics are Dice, HD95, and average surface distance. Calibration is measured with ECE both globally and inside a band within a few voxels of the reference boundary. Error detection is measured as the area under the ROC curve of each ambiguity signal for predicting misclassified voxels.

Three checks decide the proposal. First, before any training changes, on held-out vendor D, does $a(x)$ detect the errors of a frozen supervised model better than that model's own entropy? Second, does the fuzzy map beat the area-matched geometric band when both are used inside the same loss? Third, does adding the loss to MiDSS lower boundary-band ECE and HD95 on vendor D without lowering Dice?

The file contains no multi-annotator dataset. Agreement with inter-observer disagreement therefore remains an untested secondary hypothesis, and we use model error as its proxy.