Modified Efficient Net of Chest X-Ray Images for Lung Disease Classification Using Transfer Learning Approach

Chest X-ray (CXR) studies can be automatically detected, and their locations can be identified using artificial intelligence in healthcare. The World Health Organization reports that lung diseases such as COVID-19, pneumonia, tuberculosis, and lung opacity contribute significantly to global mortality. The overlapping symptoms of these diseases make accurate identification difficult. To accelerate the process and enable early detection, machine learning (ML) and traditional approaches are combined to improve disease detection. CXR images are used in this study to classify lung disease using ML and transfer learning (TL). The study primarily consists of three parts: firstly, the data augmentation technique addresses class imbalance; secondly, the image enhancement and pre-processing technique improves image quality; and finally, the TL approach EfficientNetB1 extracts features and uses weighted binary cross-entropy loss to handle a large number of false positive cases. Fine-tuning hyperparameters significantly enhance performance. AUC-PR, ROC-AUC, F1-score, and precision are among the metrics that demonstrate the proposed method to be more accurate and effective than other TL models. KEYWORDS Computer vision, COVID-19, deep learning, efficientNetB1, pneumonia, transfer learning