A DEEP LEARNING-BASED HYBRID MODEL FOR AUTOMATED LUNG DISEASE DETECTION: ADDRESSING COVID-19 DIAGNOSTIC CHALLENGES

Lung abnormalities in the post-COVID era are a common issue, demanding accurate and timely diagnosis. However, detecting these irregularities with deep learning models has its own difficulties like class imbalance, overfitting, and restricted generalizability due to heterogeneous datasets, all of which can impede accurate detection. Here we apply deep learning classification on extensive datasets compiled during and after the COVID-19 pandemic, with an ability for high-performance AI algorithms to peruse normal lungs and lungs affected by pneumonia, cardiomegaly, and COVID-19. We implemented an integrated framework of two convolution neural network architectures, Places365 GoogLeNet and EfficientNetB0. The fusion of these models employed the AdaBoost ensemble method, significantly enhancing classification accuracy. Places 365 GoogLeNet achieved a validation accuracy of 90.49%, while EfficientNetB0 reached 94.70%. By integrating these models, the classification accuracy improved to 97.48%, showcasing the effectiveness of model fusion in achieving superior performance. This framework demonstrates promise for diagnosing complex lung conditions, particularly those related to COVID-19, offering potential as a robust diagnostic tool.