A Hybrid Ensemble Approach for Skin Cancer Classification Using Deep Learning Techniques

Abstract

Detection of skin lesions is an important task in dermatology, as early diagnosis is fundamental to prevent the appearance and treatment of skin cancer. This thesis describes HySkinDetect, a hybrid ensemble model to enhance the efficacy and efficiency of classifying skin lesions. With the synergy of these two high performingdeep learning architectures (ResNet50 and DenseNet121), HySkinDetect takes advantage of those & get extra out performance. The model is capable of capturing hierarchical structures and complex features more efficiently by using the residual connection of ResNet50 and the dense connectivity of DenseNet121, which leads to a better performance in skin lesion detection. The model is tested on both the training and test data, where HySkinDetect achieves a test accuracy of 86.57%, which are much better than other aforementioned individual models such as ResNet50 (76.97%), VGG16 (74.35%), MobileNetV2 (61.23%) and DenseNet121 (72.02%). The model also exhibits good precision, recall and F1 scores vital to reliable prediction in clinical context. These findings indicate that HySkinDetect could also help in diagnostic process of skin lesions for health care providers, namely in the scope of skin cancer. Moreover, this dissertation presents the deployment of HySkinDetect in a cloud-based software web application while enhancing the model for real time execution, privacy concern and for usability by healthcare provider. The proposed system aims to facilitate fast, accurate, and interpretable analysis of skin lesions for more efficient diagnosis. Potential future work is to enhance the model’s generalization (i.e. adaptivlearning) and apply it to other fields of medical image. By providing a solid, scalable solution, HySkinDetect can make an important difference in clinical decision support by making the detection of skin lesions more accurate and faster – which in turn can help improve patient outcomes.