Class-Specific Synthetic Data Augmentation Using EA -GAN for Enhanced Skin Cancer Detection

The main difficulty in developing accurate and robust models of machine learning for diagnosing skin cancer has always been that of class imbalance. This paper hereby presents an Enhanced Attention-Driven Generative Adversarial Networks (EA-GANs) framework especially designed to address class imbalance problems related to datasets used in skin cancer. Unlike the traditional Generative Adversarial Networks (GANs), in the proposed EA-GAN, adaptive attention mechanism along with a progressive feature refinement module allows for the creation of class-specific high-quality synthetic dermoscopic images. In this proposed method, the use of the adaptive attention mechanism enhances the focus synthesis of fine-grained features of the lesions and quality of the produced images. It progressively learns to improve upon underrepresented characteristics of each class. Data generation tasks are evaluated on the ISIC 2020 skin lesion dataset. The computations of Fréchet Inception Distance and Structural Similarity Index have also been used for the evaluation of generated images to determine the realism and diversity of outputs. EfficientNet-B3 classifiers is employed to classify the skin lesion based on the expansion of data, which reached a 94.2% accuracy rate on 23% advancement of baselines for the detection methodology of the minority class. This approach, besides improving the impacts of class imbalance, improves the generalization abilities of deep learning classifiers and therefore is a scalable solution, appropriate for practical clinical use. The proposed EA-GAN model sets a new benchmark for data augmentation by GANs in the context of medical imaging: ensuring reliable, high-quality synthetic samples while addressing the constraints associated with existing datasets.