Mitigating Skin Tone Bias in Lesion Screening via Confidence-Based Routing

We built and measured a two-stage skin-lesion screening pipeline with a skin-type-dependent confidence gate and Grad-CAM explanations. Retraining raised validation accuracy from 80.84% to 88.69% and held-out accuracy for every skin type, with gains for types I–V that exceeded seed variation, but the gap between best and worst skin type grew from 7.7 to 20.3 points, the type VI gain was not distinguishable from noise, and balanced accuracy on ISIC 2019 images outside HAM10000 was 55.70% against 80.50% on the HAM10000 validation split. The gate, pre-filter and explanations work end to end, but their value for safety is unproven. Future work (not completed): split HAM10000 by lesion; compute calibration error for all sets; evaluate the gate with risk–coverage curves; ablate the v3 changes and add a control with skin-type-unbalanced Fitzpatrick17k data; collect more type V–VI images; test on phone photographs; and have der-matologists review the tier mapping and the system’s outputs.

[1]     A. Esteva et al., “Dermatologist-level classification of skin cancer with deep neural networks,” Nature, vol. 542, pp. 115–118, 2017. [2]     M. Groh et al., “Evaluating deep neural networks trained on clinical images in dermatology with the Fitzpatrick 17k dataset,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops (CVPRW), 2021, pp. 1820–1828. [3]     R. Daneshjou et al., “Disparities in dermatology AI performance on a diverse, curated clinical image set,” Sci. Adv., vol. 8, no. 32, Art. no. eabq6147, 2022. [4]     R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and A.    Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV), 2017, pp. 618–626. [5]     Y. Geifman and R. El-Yaniv, “Selective classification for deep neural networks,” in Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. [6]     P. Tschandl, C. Rosendahl, and H. Kittler, “The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions,” Sci. Data, vol. 5, Art. no. 180161, 2018. [7]     N. Codella et al., “Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the International Skin Imaging Collaboration (ISIC),” arXiv:1902.03368, 2019. [8]     M. Combalia et al., “BCN20000: Dermoscopic lesions in the wild,” arXiv:1908.02288, 2019. [9]     DermNet NZ, “Dermatology image dataset,” https://dermnetnz.org/ dermatology-image-dataset. Accessed: Aug. 18, 2026. [10]    C. Vadivelraju, K. E. S., V. M. B. Kumar, P. Suraj, and M. H. Reddy, “DermaScan AI: Deep learning system for preliminary diagnosis of dermatological manifestations,” Int. J. Creative Research Thoughts (IJCRT), 2025, paper IJCRT2505389.