A Comparative Study of Deep Learning Models in Detection of Lung Opacity

Abstract

Lung opacity detection is vital in diagnosing various pulmonary conditions, such as pneumonia, tuberculosis, and COVID-19. This study examines the efficacy of three deep learning models Visual Geometry Group 16 (VGG16), ResNet50, and Dense Convolutional Network121 (DenseNet121) in identifying lung opacity from chest X-ray images. Using datasets sourced from the COVID-19 radiography dataset and Kaggle, these models were trained and evaluated to determine their performance. The data underwent preprocessing to improve model accuracy and generalizability, including resizing, normalization, and augmentation. Each model was assessed based on classification accuracy, demonstrating distinct levels of effectiveness. DenseNet121 outperformed its counterparts with an accuracy of 91.90%, attributed to its dense connectivity that enhances gradient flow and feature reuse. VGG16 achieved an accuracy of 89.99%, benefiting from its simplicity and structured design. ResNet50, despite its more profound architecture and skip connections, lagged with an accuracy of 75.79%, likely due to challenges in capturing subtle patterns within the data. The results establish DenseNet121 as the most suitable model for lung opacity detection, offering a reliable and efficient solution for clinical applications. Its high accuracy suggests potential integration into diagnostic workflows, where quick and precise detection is critical for improving patient outcomes. Future work includes utilizing larger datasets, implementing ensemble techniques, and integrating explainable deep learning (DL) methods to improve model interpretability and reliability, promoting adopting deep learning solutions in medical imaging.