Hybrid ResNet-50 with Gradient-weighted Histogram Technique for Automated Skin Cancer Classification

Abstract

To develop an accurate and efficient automated system for skin cancer classification based on dermoscopic images and to improve diagnostic quality compared with traditional deep learning classifiers. Dermoscopic images are resized and normalized, then fed into the pre-trained ResNet-50 backbone to extract deep features. The extracted features are then fed into a hybrid classifier with fully connected layers. In addition, a GHT-based feature importance mechanism is implemented to highlight important features and attenuate the influence of less useful ones. The proposed ResGrad Hybrid model achieves a classification accuracy of 89.4%, higher than that of the ResNet-50 model (87%). These results indicate superior performance for skin cancer detection. Here, we provide a hybrid ResGrad approach that combines deep feature extraction with a feature-importance mechanism (derived from GHT) and a feature-selection method, further improving classification performance compared with the standard ResNet-50 approach.