Dermatologists and pathologists cannot afford to trace every lesion border by hand. The cheap alternative is to have them draw a few quick strokes inside and outside the lesion, called scribbles, and let an algorithm fill in the rest. That filling-in step usually decides whether two neighbouring pixels belong together by comparing their raw red-green-blue (RGB) values. Human vision does not work that way. In a perceptually uniform color space such as CIELAB, a coordinate system warped so that equal numerical distances look equally different to a person, the distance between two colors (written $\Delta E$) tracks what an observer can actually tell apart.
This note proposes that $\Delta E$ measured across a lesion border does more than improve label filling. Before any model is trained, it predicts which stretches of the border are genuinely ambiguous: places where both the algorithm and the experts themselves will disagree. If that holds, a fixed budget of expert seconds should go to scribbles everywhere plus a handful of boundary clicks placed only on low-contrast arcs. Training should also treat those arcs as uncertain rather than forcing the network to copy one annotator's guess. In short, perceptual contrast becomes a tool for allocating expert attention, not just a better pixel kernel.
Core Claim
The obvious idea at this intersection is to swap RGB for CIELAB inside the conditional random field (CRF), a smoothing model that encourages similar-looking neighbours to share a label, or inside the random-walk step of scribble pipelines. The research file lists this as an open gap. It is also a one-line preprocessing change that any reviewer would run first, so we treat it as an ablation, not a contribution.
Our claim is stronger and falsifiable. Along any lesion border, the local perceptual contrast between the inner and outer bands is a model-free, annotation-free signal. Low contrast should coincide with three things: the arcs where scribble propagation leaks or stalls, the arcs where expert annotators place the boundary differently, and, as a secondary hypothesis, images with particular perilesional skin lightness and hue. The contrast is measured as CIEDE2000 $\Delta E$ after normalizing to the surrounding skin. If perceptual contrast predicts expert disagreement better than RGB gradient magnitude, luminance-only contrast, or the network's own uncertainty, then it justifies two design changes. The first is budget-matched active boundary queries. The second is ambiguity-weighted boundary losses. The comparison against luminance-only contrast matters because it tests whether chroma carries information that grayscale training throws away.
Positioning Against Prior Work
The closest existing work is scribble-supervised salient object detection by Zhang et al.. It expands scribbles with a DenseCRF and aligns predictions to RGB image gradients through a gated structure-aware loss, reaching an ECSSD F-measure of 0.865. It works on high-contrast natural photos. It treats every edge as equally trustworthy and never asks where human labels are themselves unreliable.
In medicine, Can et al. expand scribbles by random walk, refine them with a CRF, and filter them with Monte Carlo dropout uncertainty, coming within 2.9 to 4.5 Dice points of full supervision on cardiac and prostate MRI. Their uncertainty gate is the model-dependent counterpart of our model-free contrast signal, and we compare the two head to head. Kim et al. build a differentiable level-set loss that replaces test-time CRFs in RGB natural scenes. We reuse that style of training-time boundary loss but weight it by perceptual ambiguity.
The color side supplies both method and motivation. Achanta et al. compute saliency as plain CIELAB distance to the mean image color, which shows that perceptual distance alone yields usable figure-ground maps. Ali et al. survey CIELAB clustering for scoring color variegation, but assume the segmentation mask already exists. Kittler et al. use perceptual clustering to suppress texture variance while keeping small chromatic defects, which is the behaviour we want inside variegated lesions. Thong et al. audit models along CIELAB lightness and hue angle. We borrow their two-axis description of skin for subgroup analysis. Our signal, however, is the contrast between lesion and skin, not absolute tone.
Two papers constrain the design. Nguyen reports a 0.808 Jaccard index on ISIC 2017 validation using grayscale training, and Abbott et al. show that destroying hue makes skin detectors robust. Both appear to contradict the idea that chroma matters. We address this directly with a luminance-only control: if $\Delta L^*$ predicts disagreement as well as full $\Delta E$, chroma is not doing the work and the claim narrows accordingly. Two further papers motivate the work. IMA++ quantifies expert disagreement: mean pairwise Dice of 0.866, 236 images below 0.5, and boundary shifts that depend on the annotation tool. That dataset is what makes the central hypothesis testable at all. Finally, the known limitation of CIELAB, its assumption of a standard illuminant, is why we normalize against perilesional skin and also compare OKLAB.
Research Program
The first component is a boundary-contrast map. For a candidate border, taken from consensus masks during analysis or from the current prediction during training, we sample short normal profiles. Each profile pairs an inner band with an outer band, and we compute CIEDE2000 $\Delta E$ between their robust means. Before that, we apply a gray-world correction anchored on an annulus of skin around the lesion, so that camera and illuminant shifts do not masquerade as contrast. Hair and bubble pixels are masked out with a simple morphological detector. The result is a contrast value per border arc, available for free on every image.
The second component is an analysis on IMA++. We take the 2,394 multi-annotated images and measure, for each arc of the STAPLE consensus border, the spread of annotator boundaries along the normal. We record the annotation tool as a covariate, because IMA++ reports tool-induced bias. We then ask how well arc contrast predicts that spread.
The third component is a training recipe. Experts draw scribbles. Scribbles are expanded with a perceptual random walk and CRF. A segmentation network is trained with a level-set boundary loss in the spirit of Kim et al., weighted per arc by a monotone function of contrast. High-contrast arcs are pulled firmly to the image edge. Low-contrast arcs receive a soft target or a boundary click. Clicks are allocated under a time budget: after an initial round, the arcs with the lowest contrast receive single expert clicks that pin the border. The budget uses a cost model in which a scribble takes the 1 to 2 seconds reported by Zhang et al.. We calibrate the cost of a click and of a full mask with a small timed pilot rather than assuming them.
Baselines
The naive reference point is a U-Net trained directly on scribbles with partial cross-entropy and no propagation. It shows how much any expansion helps at all.
The strongest existing scribble method is the Can et al. pipeline (random walk, CRF, Monte Carlo dropout gating) reimplemented for dermoscopy with RGB kernels. For context, its reported gap to full supervision is 88.0% versus 91.2% Dice on ACDC. We found no reported result on Lacuna for scribble-supervised dermoscopic segmentation, so we report our own reimplementation. The scribble-plus-DenseCRF system of Zhang et al., including its gated structure-aware loss, is a second strong scribble baseline.
The strongest label-efficient competitor that does not use scribbles is transformation-consistent self-ensembling from Li et al.. It reports a Jaccard index of 0.798 on ISIC 2017 with 300 labeled images, against a fully supervised 0.765 in their setup. We convert 300 dense masks into expert seconds with the calibrated cost model and match our scribble-plus-click budget to it. Beating this baseline means matching or exceeding 0.798 Jaccard at equal or lower expert time, on the same backbone.
The key ablation is the same pipeline with RGB kernels replaced by CIELAB kernels and no contrast-based routing or weighting. This is the "obvious idea." Gains beyond it isolate the effect of using contrast to allocate effort. A second ablation replaces $\Delta E$ with $\Delta L^*$ in routing and weighting, which isolates chroma.
All methods share one backbone, the same data splits, an identical augmentation set, and a fixed hyperparameter search of equal trials on a held-out validation split. Fully supervised training on all 2,000 ISIC 2017 masks is reported as a ceiling. Annotator-to-annotator agreement on IMA++ serves as the human reference.
Why not just…?
Why not just run the existing CRF on CIELAB-converted images? It may recover much of the pixel-level gain, which is exactly why it is our main ablation and not our contribution. It does not tell experts where to spend effort, and it does not stop the network from being trained on arbitrary boundaries in ambiguous regions. If routing and weighting add nothing over this ablation at matched budget, the proposal reduces to "use Lab," and we will report that.
Why not just route expert clicks by the network's own uncertainty? This is the real threat. Monte Carlo dropout already gates pseudo-labels in Can et al.. Network uncertainty, though, depends on a partly trained, possibly biased model, and it is unavailable before training. Perceptual contrast is fixed by the image. The IMA++ analysis compares the two directly as predictors of annotator spread, and the budget experiment compares them as routing rules. If uncertainty wins or ties, we test their combination. If the combination does not help, we conclude contrast is redundant.
Why not just train on grayscale with color augmentation? Nguyen reports 0.808 Jaccard this way, and Abbott et al. show that hue destruction adds robustness. These results concern fully supervised training, where dense masks already define the border. Under scribbles, the border must come from somewhere, and pigmented-on-erythematous borders can be nearly equal in luminance. The $\Delta L^*$ ablation and a grayscale-input arm of every baseline test this directly.
Why not just spend the budget on more scribbles or a few dense masks? Scribbles cover lesion interiors, not borders, so more of them should saturate. The budget curves include "all scribbles" and "a few dense masks" arms at every budget point, so this alternative is measured, not assumed away.
Proposed Experiments
The first decisive check runs on IMA++. Using the multi-annotated images, we compute, for each consensus-border arc, whether its spread exceeds the median. We then measure how well that is predicted by area under the ROC curve and by Spearman correlation for each candidate signal: CIEDE2000 contrast, OKLAB contrast, Euclidean RGB difference, image gradient magnitude, $\Delta L^*$, and Monte Carlo dropout variance from a scribble-trained model. We stratify by annotation tool. The hypothesis fails if perceptual contrast does not beat RGB difference and gradient magnitude by a clear margin under bootstrap confidence intervals.
The second check is budget curves on ISIC 2017, with ISIC 2018 five-fold cross-validation and PH2 as an external test. Scribbles and clicks are simulated from expert masks, with randomness over stroke placement. We plot Jaccard and Dice against expert seconds for every baseline and ablation. We also report boundary error near the border and a within-human-range rate. That rate is computed on IMA++ images held out from training: it is the fraction of images where the model's Dice to each annotator is at least the annotators' mutual Dice, which avoids rewarding the model for mimicking one person's tracing. A timed annotation pilot with a few volunteer clinicians checks that the simulated cost model is not wildly off.
The third check stratifies all errors by perilesional skin lightness and hue angle, following Thong et al.. It asks whether low lesion-skin contrast clusters in particular skin groups and whether contrast-based routing narrows or widens the gap compared with RGB pipelines. ISIC collections underrepresent dark skin, so we report subgroup sizes and confidence intervals. Where groups are too small to support claims, we say so rather than claim fairness gains.
References
- Weakly-Supervised Salient Object Detection via Scribble Annotations
- Learning to Segment Medical Images with Scribble-Supervision Alone
- Weakly Supervised Semantic Segmentation Using Color Adjacency Loss
- Frequency-tuned salient region detection
- Automating the ABCD Rule for Melanoma Detection: A Survey
- Detection of Defects in Colour Texture Surfaces
- Beyond Skin Tone: A Multidimensional Measure of Apparent Skin Color
- ISIC 2017 Skin Lesion Segmentation Using Deep Encoder-Decoder Network
- Color Invariant Skin Segmentation
- IMA++: ISIC Archive Multi-Annotator Dermoscopic Skin Lesion Segmentation Dataset
- Semi-supervised Skin Lesion Segmentation via Transformation Consistent Self-ensembling Model