cs.CVNov 1, 2026
SaveTesting concept-level audits of melanoma classifiers with diffusion counterfactuals and matched positive controls
Organizations: Department of Computer Science & Engineering, Rajshahi University of Engineering & Technology, Rajshahi, 6204, Bangladesh · Department of Computer Science & Engineering, American International University-Bangladesh (AIUB), Dhaka, Bangladesh · ELITE Research Lab, Queens, NY, 11435, USA · Centre for Intelligent Cloud Computing (CICC), COE of Advanced Cloud, Faculty of Information Science & Technology, Multimedia University, Jalan Ayer Keroh Lama, Bukit Beruang, Melaka, 75450, Malaysia
Abstract
Concept-based explanations of melanoma classifiers are rarely tested against controls with known targets. We built a counterfactual audit from a diffusion model conditioned on seven 7-point checklist concepts and tested it with such controls. Colour-calibrated start noise reduced colour outliers from 78.5-84.8% to 6.6-13.3%. Four of seven knobs were target-dominant but also changed other concepts. Quartile-based positive controls were confounded, and the audit failed its pre-specified rule. It recovered the targets of all 16 controls matched on the generator's pseudo-labels and of 12 of 12 controls labelled by a second evaluator, for three concepts; concept activation vectors recovered fewer (11 of 16). In 18 permuted-label classifiers, significant sweep responses and target rankings also arose, whereas standardised effect margins separated the controls from these null classifiers. The pigmentation sweep raised the melanoma logit in 15 of 15 ISIC 2020 classifiers (+3.12 on average) and 6 of 6 HAM10000 classifiers (+4.01), about four times the null response in scale-free units; the regression effect did not transfer. An exploratory decomposition, evaluated against the controls, suggested that knob cross-talk partly hides vascular responses. Matched controls and null classifiers are practical checks of concept audits.
Figures & tables
Figure 1: Study design. a , The counterfactual instrument and the audit statistic. CBM-A, the concept bottleneck model that produced the conditioning pseudo-labels; CBM-B, an evaluator trained separately on the Derm7pt expert labels. b , The validity rules in the order in which they were fixed, with their outcomes. Each rule was written into the code before the run that tested it. c , The three other methods, tested on the same controls, and the melanoma classifiers. PN, pigment network; VS, vascular structures; Pig, pigmentation; Reg, regression structures; BWV, blue-whitish veil; Str, streaks; D/G, dots and globules; BH, Benjamini-Hochberg.
Figure 2: The counterfactual instrument. a , AI-generated images (our concept-conditioned diffusion model; research outputs of this study, not clinical photographs) for one realistic benign base (the pseudo-label vector of a real ISIC 2020 benign image) with each of the seven concept knobs set from 0 to 1. Along each row the other six concepts, the start noise and the colour target do not change. Bold labels mark the four target-dominant knobs; grey labels mark the other three. b , Colour outliers (images whose mean CIELAB lies outside the 99% region of real images) for the original start noise, a trailing-timestep start and the colour-calibrated start, with the same 256 start noises per set. Dotted lines, real images; dashed line, the 20% limit fixed before the test. c , Mean change of each CBM-B concept output, in SD of real test images, when each knob goes from 0 to 1 (mean of the two realistic bases, 64 trajectories each). Outlined cells, the knob’s own concept. Right, the number of the four bases in which the knob raised its own output more than any other knob did (target-dominant).
Figure 3: Positive controls of the audit. a , b , Standardised mean difference between the high and low groups of each positive-control design, for the target concept (diamond) and the six other concepts (circles), read with the CBM-A pseudo-labels ( a ) and with the CBM-B outputs ( b ). Orange, quartile designs; blue, pairs matched on CBM-A labels; green, pairs matched on CBM-B labels. Each matched design is balanced in its own label set but not fully in the other. Labels name the most unbalanced other concept when its difference exceeds 0.3 SD. n.t., not testable (fewer than 50 training pairs). c , Change in the control classifier’s logit from knob level 0 to 1 (mean of the two realistic bases) for the 3 original controls and the 16 CBM-A matched controls. d , The same for the CBM-B matched controls, with the decision of rule 4 for each concept (valid or not valid). Outlined cells, target knob. Right of each panel, the number of bases (of 4) in which the target knob gave the largest positive and significant effect; neg., negative control.
| Method | Pre-specified rule, quartile controls | Matched controls passed, CBM-A labels | Matched controls passed, CBM-B labels | Concepts passing | Negative- control ratio (limit 0.25): logit (scale-free) | , same architecture, other seed | , different classifier |
|---|---|---|---|---|---|---|---|
| Counterfactual audit | Not valid: veil control 0 of 4 bases | 16 of 16 | 12 of 12 (Reg not testable) | PN, VS, Pig with both label sets; Reg with CBM-A labels only | 0.023 (0.47) | 0.883 | 0.717 |
| TCAV, our implementation | Not valid: negative control | 11 of 16 | Not tested | PN, VS, Reg with CBM-A labels; not Pig | 0.834 (0.834) | 0.318 | 0.072 |
| TCAV, Captum 0.9.0 | Not valid: both positives and the negative | 0 of 3 | Not tested | Not tested | 0.889 (0.889) | 0.209 | 0.252 |
| Observational attribution | No rule fixed in advance | 16 of 16 (largest coefficient) | 12 of 12 (circular) | One setting only | 0.025 (0.45) | 0.976 | 0.647 |
Table 1: Validity and reliability of four concept-attribution methods on the same classifiers. Negative-control ratio: largest absolute effect of the permuted-label classifier divided by the median of the same value for the melanoma classifiers, on the logit scale fixed in rule 1 and, in brackets, after dividing each classifier’s effects by the SD of its logits on the real test images (post hoc); TCAV scores do not depend on output scale. : mean Spearman correlation of concept profiles between classifier pairs. Captum TCAV was run on 13 classifiers, including the first three matched controls. CBM-B labels: 12 controls for pigment network, vascular structures and pigmentation; regression had too few pairs to test. For the observational method the CBM-B test is circular, because its predictors are the CBM-B outputs that define the control labels. PN, pigment network; VS, vascular structures; Pig, pigmentation; Reg, regression structures.
Figure 4: Four methods against the same controls. a , Share of the matched positive controls in which each method found the target concept (counts above the bars). Light blue, the audit of the controls labelled by CBM-B; the other bars use the CBM-A controls. For the observational method, the target had the largest coefficient; it has no significance test. n.t., not tested. Dashed line, the per-concept rule (3 of 4). b , Largest effect of the permuted-label control divided by the median for the melanoma classifiers, on the logit scale of rule 1 (solid) and after dividing by each classifier’s logit SD on the real test images (hatched, post hoc); TCAV scores do not depend on output scale. Dashed line, the limit of 0.25. c , Mean Spearman correlation of concept profiles between classifiers of the same architecture with another seed, and between different classifiers.
Figure 5: Counterfactual audit of the melanoma classifiers. a , Change in the melanoma logit from knob level 0 to 1 (mean of the two realistic bases, 64 trajectories each) for the six ISIC 2020 classifiers and the Derm7pt classifier. †, knob not target-dominant (Fig. 2 c). b , The four target-dominant knobs. Points, individual classifiers; black lines, mean of the six ISIC 2020 classifiers; grey and light-blue bars, range of the four matched positive controls with CBM-A and with CBM-B labels for the same knob; logit scales differ between classifiers, so the bars are a rough guide. ‡, audit passed with CBM-A-labelled controls only.
Figure 6: Null distribution, retraining and a second training dataset. a , Scale-free effect of each target-dominant knob (logit change from level 0 to 1, mean of the two realistic bases, divided by the SD of the classifier’s logits on the 969 real ISIC 2020 test images). Grey, null classifiers: 18 trained on permuted melanoma labels (dark) and 300 random linear heads on ImageNet features (light); dashed, 95th percentile of the permuted-label classifiers (rule 5). b , Logit effects of the 15 ISIC 2020 melanoma classifiers (seeds 1 and 2, dark blue; seeds 3 to 5, light blue; rule 6) and of the six HAM10000 classifiers (green; rule 7); black lines, group means. c , Standardised target margin (target effect minus the largest other effect, divided by the logit SD; the smaller of the two realistic bases) of the matched controls and of the permuted-label classifiers scored for the same target; dashed, largest null value. Post hoc. PN, pigment network; VS, vascular structures; Pig, pigmentation; Reg, regression structures.
Figure 7: Decomposition of the sweep effects (post hoc). A ridge regression relates the logit change of every image in the realistic sweeps to the changes in the seven CBM-B outputs and the mean CIELAB colour. a , Matched controls: mean CBM-B coefficient of the four controls of each group, scaled by the largest absolute coefficient of each control; outlined, target concept; n/4, controls with the target as largest coefficient and a bootstrap interval above 0. b , c , Parts of the pigmentation-sweep and vascular-sweep effects in the six ISIC 2020 classifiers: coefficient mean measured change at level 1 (both realistic bases). Positive parts stack to the right and negative parts to the left of 0; grey, the part not explained by the measured changes; diamond, measured sweep effect.
More in this topic
Explainable Medical Image Analysis
Explainable Medical Image Analysis develops methods that make medical image predictions understandable, verifiable, and useful for clinical review.
See all 262 papers