Attention-Based Hybrid RGB-Edge-Histogram Network for Melanoma Classification

Abstract

Even with the photos of skin lesions, the identification of the melanoma remains challenging, because lesions that appear to be benign could have similar colour, texture, and structural features to melanoma. In this paper, we suggest a hybrid DL system to automatically classify melanomas with the help of complementary picture representations. For the implementation, a data set of 1001 labelled images of the benign and malignant cases is used. The dataset is randomly undersampled to get a balanced set of 1000 images. The photographs are scaled to 50×50 pixels and rendered in three different forms: a normalised RGB image, a Canny edge map and an intensity histogram of the image with 256 bins. These representations are then fed into special convolutional and dense branches that learn features, which are then concatenated and further enhanced by layer norm and four-head multi-head self-attention before being classified using binary softmax. The framework is trained using the Adam optimiser with categorical cross-entropy, and is evaluated against a CNN baseline using only images. The proposed hybrid model is able to achieve best validation accuracy of 86.50%, while image-only baseline model can achieve the test accuracy of 88.06% in the same experimental setting. The results show that it is possible to incorporate appearance, structural and intensity-distribution information into an attention-based classification pipeline, but also emphasise the necessity for bigger and externally validated datasets to provide robust generalisation. Furthermore, the solution also allows batch and real-time inference, suggesting the potential use as a computational screening aid.