Skin lesion sythesis and classification using an improved DCGAN classifier

This study addresses the challenge of early skin cancer detection, which is crucial for improving patient prognosis. Despite advances in medical imaging, many cases are diagnosed at advanced stages, limiting treatment effectiveness. Automated diagnostic technologies, such as dermoscopy, can facilitate early detection by identifying skin lesions. However, the development of reliable automated methods is hindered by limited annotated datasets and class imbalances. Deep learning models have shown promise in medical image analysis but typically require large amounts of labelled data for effective training. To overcome these challenges, this research proposes a skin lesion synthesis and classification framework based on an improved Deep Convolutional Generative Adversarial Network (DCGAN). The system generates realistic synthetic images using convolutional neural networks, while preprocessing techniques—including scaling, normalization, sharpening, colour transformation, and median filtering—enhance image quality during training. The model incorporates optimized generator and discriminator networks, global average pooling with fractional strides, and backpropagation with an adjusted learning rate to improve training efficiency and image realism. For classification, the discriminator’s final layer is adapted to predict the target class. The framework is evaluated on the ISIC2017 dataset for binary classification of benign versus malignant lesions. Performance metrics include accuracy, recall, precision, F1-score, and Balanced Accuracy Score (BAS) to account for dataset imbalance. The proposed DCGAN-based classifier demonstrates high performance, achieving superior results across all evaluation metrics compared to existing deep learning approaches. These findings suggest that the model effectively generates high-quality synthetic images and accurately classifies skin lesions, highlighting its potential as a robust tool for deep learning-driven medical image analysis.