Enhancing Skin Disease Classification Through GAN-Generated Synthetic Images for Improved CNN Training and Generalization

The work in this project helps in improving the classification of skin diseases using the combination of Generative Adversarial Networks (GANs) and Convolutional Neural Networks (CNNs). GANs were used to generate synthetic, realistic images of the rare underrepresented skin lesion types to address the problem of lack of labeled data and class imbalance. When trained on the CNN model, the model performed well, especially with a combination of experiments on real vs synthetic data, and was better able to provide improved accuracy and generalization for rare conditions like actinic keratosis and vascular lesions. Other than this, other advanced techniques like WeightedRandomSampler and especially mixed precision training provided other performance gains, by better balancing the class representation as well as speeding up the training. To detect the skin disease in clinical settings, imp and Flask have been used to deploy the skin disease classification model in real time. Through this solution we only integrate synthetic data, balanced sampling and real time inference to offer a simple, robust and scalable solution for early and accurate skin disease diagnosis. The central idea of the project is to use the advantage of GANs with CNNs to enhance the already existing classification of skin diseases. Since the authors work on a rare skin lesion diagnosis in which annotated images are scarce, availability of training data marked by class imbalance was a major avenue of concern to them. The project applies GANs to produce new, but realistic, images of underrepresented skin lesion types thereby broadening the training data set and handles this problem. Moreover, elaborate experiments concerning CNN's behaviour with both real and synthetic data are presented as additional research. Findings also demonstrate that it is accurate and capable of fulfilling the criterion of a good diagnostic tool, and further applying it to the cases of the newly discovered diseases such as actinic keratosis and vascular lesions. The features utilized to ensure the balance between the classes of the dataset and to increase the training speed, but the accuracy improvement is retained, are the use of features such as WeightedRandomSampler and mixed precision training. Thus, it is utilized as a real-time source using streamlit and Flask to generate real world inferences for clinic use skin disease classification models. Last but not least, considering the efficient synthesization of data corrupted because of balanced sampling techniques and nimble deployment of models in a friendly environment, this project provides a sound and scalable model for an early and accurate skin disease diagnosis thereby improving patients' experience in the field of dermatology.