MelaFuse-Net: Cross-Attentive Hybrid Deep Learning for Dermoscopic Melanoma Classification under Lesion-Disjoint Evaluation

Abstract

Melanoma accounts for a small fraction of skin malignancies yet for most of the deaths they cause, so a dermoscopic classifier is clinically useful only when it separates melanoma from the benign pigmented lesions it imitates. This work presents MelaFuse-Net, a hybrid network that treats a dermoscopic image as 3 complementary evidence streams and reconciles them by attention. 2 frozen ImageNet encoders, EfficientNet-B0 and ResNet-50, convert the preprocessed image into 2 grids of 49 tokens whose channels are whitened to a common width. Prior-biased co-attention blocks then let every token of one stream query the tokens of the other, while a lesion prior obtained by unsupervised segmentation adds a learnable per-head bias to each attention logit, so that query capacity is spent on lesion tissue rather than on surrounding skin. A parallel branch computes 44 asymmetry, border, colour, geometry and texture descriptors and turns them into a gate that rescales the visual code before classification. Only 1.08 million parameters are trained. On the HAM10000 collection of 10,015 dermoscopic images, evaluated with 5-fold cross validation in which no lesion contributes images to more than one fold, MelaFuse-Net reaches 74.13% accuracy and 55.59% macro F1, the best of 7 models measured under an identical protocol, and exceeds the single-stream token transformer by 5.85 accuracy points.