Multi-Scale Multi-Class Generative Adversarial Network for Improving Deep Learner based Lung Disease Detection

Objective: The prediction results of Deep Learning (DL) model rely heavily on labeled training images and the quality of images in the datasets. The most of datasets consists of blurry and low quality images due to reconstruction and compressions processes. The main objective of this research is to generate quality and sufficient images for improving lung disease prediction of CovLscan. Methods: Radiopaedia-Chest X-Ray and Kaggle-COVID-19 Lung CT datasets are used for lung disease prediction. Multi-Scale Multi-Class Generative Adversarial Network (MsMcGAN) model is proposed for generating quality images to improve lung disease prediction accuracy of CovLscan. MsMcGAN uses Multi-Level Dilated Residual Blocks (ML-DRBs) to generate lung images that preserves visual structure and contextual disease information. A Patch- based Multi-scale Discriminator (PMsD) is created for fine-tuning the generator through distinguishing the patches representing varying scales of lesions of generated images from real images. Perceptual content (PC) loss is added to map between produced and reference modalities at varying levels of semantic abstraction. MsMcGAN model enhances lung disease prediction of CovLscan by increasing labeled lung disease images through generating high-quality target images. Findings: The test outcomes reveal that the MsMcGAN - CovLscan model accomplishes an overall accuracy of 97.5% and 97.6% on the collected CXR and CT images that is contrasted with the classical CNN models. Novelty: This study uniquely integrates ML-DRBs with a PMsD to enhance image synthesis and lung disease detection, combining multi-scale contextual feature extraction and adversarial training stability, which has not been explored in previous studies. Keywords: Lung disease Classification, Generative Adversarial Network, Chest X-Ray, Computed Tomography, Deep Learning