SyntDiaNet: Integrating feature extraction, transfer learning and classifier-embedded generative adversarial network for advanced pneumonia diagnosis

Pneumonia, a significant contributor to global disease and mortality, imposes a substantial burden on healthcare systems. Conventional approaches to pneumonia detection rely on constrained datasets and traditional feature extraction techniques, potentially resulting in suboptimal accuracy. To overcome these limitations, we introduce the "SyntDiaNet" framework, an innovative solution that amalgamates synthetic data generation and advanced diagnostic capabilities. This research presents a comprehensive methodology aimed at enhancing pneumonia detection precision. The central innovation revolves around the integration of a Classifier-Embedded Generative Adversarial Network (CE-GAN) for data augmentation, followed by a multifaceted feature extraction process. This process encompasses Contour Detection, Gabor Texture Analysis, and Blob Detection, to capture nuanced patterns that may indicate pneumonia in X-ray images. Leveraging broad knowledge, an unsupervised feature extraction step is integrated using the SyntDiaNet Transfer Learning model. Across diverse feature extraction strategies, combined with the CE-GAN dataset and SyntDiaNet model, accuracy rates demonstrate substantial improvements, with potential ranges of 90.5% to 97.6%. The precision ranged between 0.932 and 0.981; recall scores spanned from 0.896 to 0.987; F1 scores varied between 0.90 and 0.980, and the AUC-ROC values ranged from 0.90 to 0.97, showcasing the ability of the models to discriminate between positive and negative classes, emphasizing differences in the classifiers' overall performance in ranking instances. By utilizing synthetic data and advanced feature extraction techniques, it adeptly addresses challenges related to limited datasets and imbalanced class distributions. Furthermore, its adaptability to various medical imaging tasks underscores its potential impact across a range of healthcare applications.