Enhancing lung nodule classification with variational autoencoder-based image augmentation

Data augmentation is crucial for improving Deep Learning (DL) model performance, particularly in applications where large datasets are unavailable. In this study, we investigate a state-of-the-art approach using Variational Autoencoders (VAEs) to enhance the performance of model on imbalanced class with small dataset. Focusing on lung nodule classification within the Lung Image Database Consortium (LIDC) dataset, we utilize VAEs to generate synthetic images, expanding the dataset with reconstructed images that improve model training. Our results show that augmenting the dataset with VAE-generated images increases test sensitivity by 3.13% in the lung nodule spiculation classification task compared to the non-augmented model. This work highlights the potential of VAE-based augmentation to improve sensitivity and the overall effectiveness of computer-aided diagnosis systems in lung nodule spiculation classification.