Melanoma Detection Using Deep Learning and Cool & Warm Color Segmentation

Abstract Melanoma is a highly aggressive form of skin cancer with a significantly improved prognosis when detected early. Traditional diagnostic methods often suffer from subjectivity and accessibility limitations, prompting the need for automated and accurate solutions. This study proposes a novel framework that integrates ResNet50 a deep convolutional neural network with a biologically inspired color segmentation technique that differentiates warm and cool color regions in dermoscopic images. The segmentation method filters out non-informative image areas, allowing the network to focus on clinically relevant zones associated with melanoma features. The model was trained and evaluated on the HAM10000 dataset, which includes over 10,000 dermoscopic images spanning seven lesion categories. Extensive data preprocessing, augmentation, and a 5-fold cross-validation strategy were applied to ensure generalizability. Quantitative results demonstrate that the proposed ResNet50+C&W model outperforms baseline models, achieving an accuracy of 90.3%, ROC-AUC of 88.0%, precision of 89.1%, recall of 88.2%, and F1-score of 88.8%. Threshold analysis identified 0.6 as the optimal decision boundary, balancing sensitivity (91.4%) and specificity (75.4%). These findings confirm that combining deep residual feature extraction with color-based lesion segmentation enhances diagnostic accuracy and interpretability. This framework offers a promising direction for scalable and clinically viable melanoma detection systems.