CAMEO: A Class-Activation-Mapped Equitable Overlay Framework for Fair and Robust Deep Learning-based Skin Condition Diagnosis

1.1 Motivation: access, trust, and skin tone

• 28.6 weeks was the median referral-to-treatment wait in Canada in 2025, and 61% of surveyed healthcare professionals named data bias as a barrier to trusting and adopting AI.The passage also reports a global average wait of 70 days.Source: §1.1 · p. 1
• On Fitzpatrick types V and VI, redness may appear purple, brown, or grey, and darker-skinned patients face missed early disease and underdiagnosis of inflammatory conditions.The cited passages connect these clinical differences and dataset representation gaps to the need for fairer diagnosis.Source: §1.1 · p. 1

1.2 From skin tone bias to shortcut learning

• Dermoscopy classifiers can learn background skin tone, device vignetting, and measurement rulers instead of lesion morphology, motivating a test of XAI-guided augmentation without accuracy loss.These shortcuts may be hidden by balanced in-distribution evaluation and appear as gaps across skin tones, devices, and clinical settings.Source: §1.2 · pp. 1–2

1.3 Overview of the proposed framework

• CAMEO asks the classifier which pixels it uses, screens its explanations for stability, builds an annotation-free lesion mask, and repaints only the surrounding skin with textured synthetic Fitzpatrick tones.The framework identifies stability screening and calibrated skin-tone replacement as key design choices.Source: §1.3 · p. 2

Fairness in dermatology AI

• CAMEO injects Fitzpatrick diversity through background replacement and uses the model's own stable explanations to select the region, closing a loop between explanation and optimization without demographic labels.The passage contrasts this with adversarial debiasing that requires sensitive-attribute labels and operates on feature vectors.Source: §2 · p. 2

Lesion segmentation

• The XAI-derived lesion mask is less precise than prompted, fine-tuned SAM, but is automatic, annotation-free, and performs better than unprompted SAM in the target setting.The passage reports IoU 0.422 versus 0.785 for prompted SAM, while unprompted SAM reaches about 0.16.Source: §2 · p. 2

Explainability methods

• Prior explainability methods include GradCAM++, Integrated Gradients, SSIM-based stability measurement, and FGSM; surveys motivate moving XAI from reporting into training.Attribution maps are model-dependent and seldom validated before use, motivating CAMEO's stability screen.Source: §2 · pp. 2–3

Bias-adversarial augmentation

• Unlike BiasAdv, which uses adversarial perturbations, CAMEO uses attribution maps and produces visually inspectable ChromaSwap images.The passage also distinguishes the methods by application domain: dermoscopy versus natural images.Source: §2 · p. 3

Domain adaptation and covariate shift

• CAMEO adapts SIMLea's R2 feature-space distance to separate external images resembling training data from out-of-distribution cases using a threshold at the 5th percentile of validation scores.The distinction is used to separate genuine classification error from domain mismatch.Source: §2 · p. 3

3.1 Research Gap

• Dermoscopy classifiers can rely on background cues, while spatially uniform augmentation leaves unperturbed regions available for shortcuts and attribution-derived masks can inherit ruler artifacts.The paper also identifies limited systematic evaluation and correction for darker skin tones as a gap.Source: §3.1 · p. 3
• Post-hoc XAI documents model bias without correcting it, and attributions used in training are rarely stability-checked despite being model-dependent and unreliable on some already-misclassified images.The paper frames this as a separation between auditing and fixing bias.Source: §3.1 · p. 3

3.2 Novelty

• CAMEO screens attribution maps under small Gaussian perturbations with SSIM before they affect training, because unstable maps may encode fragile shortcuts and reinforce bias when used for repainting.The stability score also determines how strongly each image is augmented.Source: §3.2 · p. 3, §3.3 · p. 4
• ChromaSwap replaces lesion backgrounds with textured Fitzpatrick I–VI skin, preserving visually identifiable skin rather than using flat or arbitrary fills that may create anomalies.The paper argues calibrated replacements keep augmented samples in-distribution or within manageable shift.Source: §3.2 · p. 4, §3.3 · p. 4

3.3 Contributions

• Across five seeds, CAMEO is compared with random-background, adversarial gradient-reversal, and BiasAdv baselines; the reported gain is attributed to background decorrelation, not a particular palette.Fitzpatrick calibration is presented as the clinically realistic instance of that mechanism, without demographic labels.Source: §3.3 · p. 4
• The composite XAI masks reach mean IoU = 0.422 against 2,332 dermatologist-drawn segmentations and IoU ≈0.16 for unprompted SAM [27], without external annotation or a foundation model.The masks are derived from the classifier being corrected and are intended to reflect the regions it actually uses.Source: §3.3 · p. 4
• An R2 feature-space filter with a percentile threshold separates genuine classification errors from domain mismatch on an external dark-skin test set.The stated aim is to avoid penalizing a model for images statistically foreign to its training distribution.Source: §3.3 · p. 4

4.1 Datasets and Notation

• HAM10000 provides binary melanoma classification data, while darker-skin ISIC images form the external test set; Base-Balanced is the primary raw-image baseline.The frozen HAM10000 test set is class-balanced, and T1 and T_swap probe distinct shifts that are not compared directly.Source: §4.1 · p. 4, §4.2 · p. 5

4.3 Proposed Framework Pipeline

• CAMEO computes attribution maps, screens their stability, suppresses ruler responses, forms lesion masks, replaces backgrounds, then retrains and evaluates.The stability screen determines which explanations may be acted on and how strongly each sample is augmented.Source: §4.3 · p. 5, Figure 1 · p. 6

4.4 Step 1: Screening Explanations for Stability

• GradCAM++ and Integrated Gradients maps are compared with perturbed-input maps using averaged SSIM; unstable samples receive stronger, broader background augmentation rather than being discarded.Mean SSIM below 0.7 flags instability; flagged samples contribute two variants from the full Fitzpatrick I–VI range, while stable samples contribute one from a narrower range.Source: §4.4 · pp. 5–6

4.5 Step 2: Ruler Artifact Suppression and Composite Mask Generation

• Ruler responses are zeroed in both attribution maps before fusion, and the corrected maps are normalized, averaged, thresholded, and morphologically refined into an annotation-free binary lesion mask.The pipeline excludes vanilla saliency and applies a 60th-percentile threshold followed by cleanup operations.Source: §4.5 · p. 7

4.6 Step 3: Semantically-Aware ChromaSwap Augmentation

• ChromaSwap keeps lesion pixels unchanged while replacing surrounding skin with textured synthetic Fitzpatrick tones, avoiding flat fills and generated lesion morphology.The replacement spans Fitzpatrick types I–VI, with feathered boundaries and textured backgrounds.Source: §4.6 · p. 7

4.7 Step 4: Covariate-Shift Detection via R2 Feature-Space Similarity

• Covariate shift is assessed in CNN feature space using PCA-subspace reconstruction R2, with a validation-derived threshold that labels at least 95% of in-domain images as in-distribution.At tap 4, the threshold is τ = 0.9426.Source: §4.7 · pp. 7–8

4.8 Evaluation Protocol

• Evaluation covers clean accuracy, overfitting, attribution stability, external dark-skin generalisation, background-counterfactual stability, and FGSM robustness; ITA is excluded as an unreliable pigmentation proxy.The XAI-derived swap masks have mean IoU ≈0.42, are frozen across models, and make T_swap a relative robustness probe rather than an absolute clinical-accuracy measure.Source: §4.8 · p. 8

4.9 Code Availability

• The authors state that implementation scripts, evaluation manifests, and the frozen T0 and T_swap test sets will be made available.The planned release includes the stability screen, ruler suppression, mask generation, ChromaSwap, and covariate-shift filter.Source: §4.9 · p. 8

5.1 Clean Performance on T0

• CAMEO achieved accuracy 0.836 [0.795, 0.874] and the highest AUC-ROC, 0.913 [0.880, 0.942], versus Base-Balanced accuracy 0.848 [0.810, 0.886] and AUC-ROC 0.912 [0.881, 0.941].The paired McNemar test found no significant accuracy difference (p = 0.62); CAMEO-Skewed improved over Base-Skewed by 5.3 pp.Source: §5.1 · pp. 8–9, Table 3 · p. 9

5.2 Overfitting Check

• CAMEO’s train-minus-test accuracy gap was 0.159, within the baseline range and smaller than CAMEO-Skewed’s; the augmentation introduced no overfitting problem of its own.CAMEO-Skewed had the largest gap (0.196), consistent with overfitting to synthetic melanoma backgrounds.Source: Table 4 · p. 9, §5.2 · p. 9

5.3 XAI Instability Analysis

• Of 3,000 extracted training images, 744 (24.8%) had unstable explanations and 1,559 (52.0%) exceeded the background-bias criterion.Flagged samples receive a stronger ChromaSwap schedule, applying additional background-decorrelation pressure where attention is fragile.Source: §5.3 · p. 9, Table 5 · p. 9

5.4 ITA Skin-Tone Proxy Validation (Unsuccessful)

• ITA-based skin-tone bins were unreliable because “Dark” images often reflected shadows, hair, lesion pigmentation, or imaging artifacts rather than skin tone.The authors therefore do not report ITA-stratified accuracy, relying instead on the external T1 set and background-counterfactual probes.Source: §5.4 · pp. 9–10

5.5.1 Covariate-Shift Quantification

• T1 had mean R2 = 0.929, Rshift = 1.751 at tap 4, and a 63% in-distribution rate, while T0 had R2 = 0.965 and a 93.3% rate.The T1 shift was about 13 times larger than T0’s; more than a third of T1 images lay outside the HAM10000-trained feature manifold.Source: Table 6 · p. 10, §5.5.1 · p. 10

5.5.2 Filtered Accuracy on T1

• Base-Balanced accuracy on T1 rose from 0.640 on the full set to 0.810 on the 63 in-distribution images; CAMEO rose from 0.660 to 0.794.Base-Balanced fell from 0.848 on T0 to 0.640 on full T1, a 20.8 pp loss; filtering recovers it to within 3.8 pp of T0.Source: Table 7 · p. 10, §5.5.2 · p. 10

5.6 Attention and Decision Stability under Background Swap

• Across five seeds, CAMEO reduced the T0-to-T_swap accuracy loss from 8.2 pp to 2.1 pp; label flips under background swaps fell from 19.8% to 5.4%.Mean probability range fell from 0.250 to 0.123, and GradCAM++ SSIM after a dark-background swap rose from 0.720 to 0.842.Source: §5.6 · pp. 10–11
• In one non-melanoma example, Base-Balanced changed its prediction to melanoma as the background darkened, with probability rising from 0.46 to 0.90, while CAMEO stayed at 0.09.The accompanying GradCAM++ comparison shows baseline attention scattering onto surrounding skin while CAMEO’s remains on the lesion.Source: §5.6 · p. 11
• On clean in-distribution images, Base-Balanced had a higher mean lesion-attention ratio than CAMEO (0.55 versus 0.44); ChromaSwap’s benefit is not sharper localization there.The reported benefit is instead greater invariance of diagnosis to background skin tone.Source: §5.6 · p. 11

5.7 XAI Mask Validation against Ground-Truth Lesion Masks

• Composite XAI masks reach mean IoU 0.422 without prompts, ground-truth mask fine-tuning, or an external segmentation model, versus about 0.16 for the without-prompt SkinSAM baseline.Across 2,332 matched images, mean Dice is 0.570; Otsu has higher raw IoU on HAM10000, but depends on global grayscale intensity.Source: §5.7 · p. 11, Table 10 · p. 12, Table 9 · p. 12
• Otsu’s higher raw IoU comes with greater per-image dispersion and more catastrophic failures: 6.8% below IoU 0.1 versus 4.1% for the composite masks.The authors caution that raw mask overlap is not the paper’s robustness lever; background replacement drives robustness largely independently of mask source.Source: §5.7 · p. 11

5.8 Exhaustive Fitzpatrick Augmentation Showcase

• Across Fitzpatrick I–VI, the same lesions retain their interiors pixel for pixel while textured synthetic backgrounds replace the surrounding skin.The examples are carried through all six levels; the augmentation is applied without selecting examples by mask quality.Source: §5.8 · p. 12

5.9 Augmentation Ablations

• 0.810 versus 0.792 for both LesionJitter and LesionJitter+B on T_swap shows background replacement, not lesion editing, drives the ablation’s robustness gain.Legacy-FlatBG reaches 0.564 on T_swap despite 0.787 on T0, indicating that a constant-colour background does not provide the same result.Source: §5.9 · p. 14
• Legacy-Textured matches CAMEO on T0 at 0.836 but scores 0.787 versus 0.810 on T_swap, consistent with warm-starting from a model that learned the background shortcut.CAMEO uses a balanced core with symmetric textured ChromaSwap from ImageNet initialisation.Source: §5.9 · pp. 14, 16

5.10 Multi-Seed Variance and Debiasing Baselines

• Across five seeds, CAMEO improves T_swap accuracy over Base-Balanced in every seed, with a mean gain of 7.0 pp and a range of 1.8 to 13.2 pp.The T0-to-T_swap accuracy drop is 0.021 ± 0.022 versus 0.082 ± 0.035, and seed-to-seed variance is roughly halved.Source: §5.10 · p. 16, Table 12 · p. 16
• CAMEO scores 0.810 and CAMEO-RandomBG 0.816 on the same 342 T_swap images; the difference is not significant (McNemar p = 0.87).The authors therefore do not claim Fitzpatrick-specific tones are necessary for the robustness gain.Source: §5.10 · p. 16, Table 13 · p. 17
• CAMEO outperforms adversarial gradient-reversal debiasing at 0.810 versus 0.734 (p < 0.01) and is tied with BiasAdv at 0.810 versus 0.798 (p = 0.62).The authors describe Fitzpatrick calibration as a clinically realistic instance of lesion-preserving background decorrelation.Source: §5.10 · p. 16
• The XAI pipeline contributes stability screening, annotation-free model-faithful localisation, and interpretability, even though robustness is largely agnostic to mask source and palette.About a quarter of the extraction set produces explanations that move under imperceptible noise; unstable maps risk writing model-specific artifacts into training data.Source: §5.10 · pp. 16–17

5.11 Qualitative Mask and Augmentation Analysis

• The cleaned masks localise lesions compactly and exclude the ruler zone, while augmented images preserve lesion structure and use realistic, diverse skin-tone backgrounds.Figure 9 presents the original image, ruler-corrected GradCAM++ heatmap, cleaned composite mask, and ChromaSwap image in sequence.Source: §5.11 · p. 17

5.12 FGSM Adversarial Robustness

• At ε = 0.01, Base-Balanced falls from clean accuracy 0.848 to 0.085, while CAMEO-Skewed reaches 0.272 despite lower clean accuracy.CAMEO improves over Base-Balanced at every FGSM epsilon up to 0.05 while retaining most of its clean accuracy; Table 14 reports adversarial accuracy on frozen T0.Source: §5.12 · p. 17, Table 14 · p. 18

6 Discussion

• A quarter of training-set explanations move under input noise too small to see, and about half concentrate saliency at the border rather than the lesion.Stability screening converts explanation reliability into per-sample augmentation strength, so low-confidence maps receive extra debiasing pressure instead of silently guiding edits.Source: §6 · p. 18
• Base-Balanced reaches T0 accuracy 0.848; CAMEO trails by 1.2 pp while matching its AUC-ROC, so ChromaSwap does not harm clean performance.The discussion identifies class balancing as the dominant in-distribution factor.Source: §6 · p. 18
• On T1, Base-Balanced’s 0.640 accuracy is largely attributed to domain mismatch: 37% of the set lies outside the trained feature manifold at 13 times T0’s shift.Restricting evaluation to the in-distribution subset recovers 0.810.Source: §6 · p. 18
• CAMEO’s decisions and attention are more invariant to background tone: label flips are 5.4% versus 19.8%, and CAM SSIM is 0.84 versus 0.72.Clean T0 accuracy does not differ significantly from Base-Balanced (McNemar p = 0.62); controls attribute the robustness gain to lesion-preserving background decorrelation rather than a specific palette.Source: §6 · p. 18
• Applying ChromaSwap only to the minority class yields a 0.196 train-minus-test gap and 0.778 accuracy on in-distribution T1, so synthetic augmentation does not replace real-data class balancing.The discussion also characterizes clean accuracy, adversarial robustness, and distribution-shift robustness as distinct axes that can trade off.Source: §6 · p. 18

7.1 Limitations

• Fitzpatrick calibration is a realistic, fairness-relevant instance of background-diversity augmentation, not the unique cause of robustness improvement.A larger, genuinely tone-stratified cohort would be needed to show an advantage for tone-specific over generic backgrounds.Source: §7.1 · pp. 18–19
• The unrun Otsu-mask-plus-random-background control leaves the XAI mask’s separate contribution to robustness unresolved.Otsu’s intensity assumption also weakens on Fitzpatrick V and VI skin, so the control needs repetition on a tone-diverse cohort.Source: §7.1 · p. 19
• Routing the 24.8% of XAI-unstable images to stronger ChromaSwap is motivated, not empirically optimized, because discard-versus-reweight was not tested.Source: §7.1 · p. 19
• Single-seed baselines and a small T1 set limit evaluation; T_swap’s predicted masks have mean IoU ≈0.42, so absolute accuracy can reflect masking error as well as background sensitivity.The identical masking noise across models preserves the fairness of relative comparisons.Source: §7.1 · p. 19
• The study uses one EfficientNet-B4 backbone and HAM10000-specific ruler-suppression settings; its training data are skewed toward lighter skin and early-stage flat lesions.The external dark-skin set and background-counterfactual probes, rather than internal ITA bins, support the fairness claims; 37% of that external set lies outside the trained manifold.Source: §7.1 · pp. 18–19

7.2 Future Work

• Future work proposes larger demographically and stage-stratified cohorts, transfer-feasibility analysis on T1, component ablations, and five-seed baseline replication.Source: §7.2 · p. 19
• The authors propose extending CAMEO to pigmentation-sensitive conditions such as pressure-injury staging, atopic-dermatitis grading, and surgical-site-infection detection.Source: §7.2 · p. 19
• They also propose testing the shortcut-correction principle in another imaging domain, such as chest X-ray or retinal fundus imaging.Source: §7.2 · p. 19

8 Conclusion

• CAMEO screens explanation stability, fuses corrected attribution maps into annotation-free lesion masks, and replaces surrounding skin with textured synthetic Fitzpatrick tones while preserving lesion pixels.Source: §8 · p. 19
• Background-swap melanoma decision flips fall from 19.8% to 5.4%, while GradCAM++ map SSIM is 0.84 versus 0.72; clean accuracy is within 1.2 percentage points of the class-balanced baseline (McNemar p = 0.62).CAMEO also achieves the highest AUC-ROC among the models and reaches 0.794 accuracy on the R2-filtered in-distribution subset of the external dark-skin set.Source: §8 · p. 19
• The gains require no demographic labels, external segmentation supervision, or additional annotation, and the skin-editing mechanism is proposed for other pigmentation-sensitive conditions and imaging domains.Source: §8 · p. 20

• The project received support from the Talent Recruitment Fund from the New Brunswick Innovation Foundation and the Harrison McCain Foundation Young Scholars Award, 2024.Source: Funding · p. 20