STAN: A Skin-Tone Adaptive Network for Multi-Branch Dermoscopic Lesion Classification

Abstract

Skin-lesion classification is affected by variation in lesion morphology, image quality, class distribution, and differences in skin appearance. This study presents STAN (Skin-Tone Adaptive Network), a multimodal architecture designed to combine complementary spatial, frequency, and clinical-shape information while conditioning feature fusion on an estimate of skin tone. The model uses a custom multi-scale convolutional network for RGB images, a single-level Haar-wavelet branch for frequency information, and a four-dimensional set of handcrafted lesion descriptors derived from the ABCDE framework. The three representations are projected to a common feature dimension and combined using an Individual Typology Angle (ITA)-conditioned softmax gate. Feature-wise Linear Modulation (FiLM) is then applied to the fused representation. An auxiliary regression head predicts the four handcrafted descriptors during training. The experiment used a five-class subset prepared from the ISIC 2020 image collection: Actinic Keratosis, Basal Cell Carcinoma, Melanoma, Normal Skin, and Squamous Cell Carcinoma. A total of 15,675 images were used. The proposed methodology created a stratified 80:20 split, giving 12,540 training images and 3,135 validation images. The reported validation performance was 90.40% accuracy, 90.32% macro precision, 90.44% macro recall, 90.37% macro F1-score, and 98.81% macro ROC-AUC. The strongest class was Normal Skin, while Actinic Keratosis was the most difficult class. The main errors involved Actinic Keratosis and Basal Cell Carcinoma. These results are encouraging for the proposed multimodal architecture, but they do not establish clinical generalization or the independent contribution of each architectural component. Further work should include an independent test set, repeated runs, ablation experiments, calibration analysis, and subgroup evaluation.