Skin Cancer Diagnosis Using Deep Convolutional Generative Adversarial Network for Image Classification

Skin cancer is a rapidly spreading and critical global health concern. Due to limited resources and the urgent need for early detection, accurate diagnosis, and preventive measures are paramount. Dermatologists continue to have difficulties with early detection, which is why deep learning methods have grown increasingly popular recently. Convolutional neural networks, or CNNs, have shown to be exceptionally successful models for tasks involving the identification and classification of objects. In this work, the MNIST: HAM10000 dataset, comprising seven distinct types of skin lesions, to address this pressing issue, with a dataset of 10,015 samples, taken for a comprehensive investigation. To enhance the data quality, the study employs various data pre-processing techniques, including sampling, sharpening using segmentation via auto-encoder and decoder architecture. Additionally, the study proposes the use of a Deep Convolutional based Generative Adversarial Network (DCGAN) as a novel approach to generate synthetic skin cancer images. This augmentation technique significantly expands our training dataset, aiding in the enhancement of classifier performance. The DCGAN generates synthetic images that closely resemble real skin lesions, facilitating more robust training for our CNN, DenseNet169, and ResNet50 models. Proposed experimental results demonstrate the effectiveness of our approach. By combining DCGAN-generated data with transfer learning, the study achieves higher accuracy in skin cancer diagnosis compared to conventional methods. Our findings offer promise for early detection, thereby contributing to more effective preventive measures and improved patient outcomes in the fight against skin cancer.