A Cloud-Based Hybrid Deep Learning Framework for Automated Skin Cancer Classification Using the ISIC Dataset

Abstract

Early identification of malignant skin lesions is critical for reducing mortality associated with skin cancer. Manual examination through dermoscopy followed by biopsy, although clinically reliable, is time-consuming and dependent on specialist availability. This research designs a hybrid neural network architecture aimed at enabling automated large-scale detection of skin cancer through multi-class analysis of dermoscopic images collected from the ISIC archive. The proposed model integrates two complementary convolutional neural network backbones— Xception and ResNet50—initialized with ImageNet weights and configured in a parallel feature extraction architecture. Preprocessed images are resized and normalized before being passed through both networks, where high-level feature representations are extracted and concatenated into a unified descriptor. The combined feature representations are subsequently passed through a series of dense layers integrated with batch normalization, weight regularization, and dropout mechanisms to improve model stability and prevent overfitting. The final classifier assigns each input image to one of four diagnostic categories: Actinic Keratosis, Basal Cell Carcinoma, Melanoma, or Squamous Cell Carcinoma. Performance assessment conducted on 1,866 validation samples indicates stable and reliable predictive outcomes across all lesion classes. A web-based interface developed using Streamlit enables real-time inference, facilitating practical deployment as a clinical decision- support tool. The results indicate that combining heterogeneous CNN architectures improves robustness and discriminative capability in automated dermatological image analysis.