Multi-Class Classification of Lung Diseases Using CNN Models

In this study, we propose a multi-class classification method by learning lung disease images with Convolutional Neural Network (CNN). As the image data for learning, U.S. National Institutes of Health (NIH) dataset divided into Normal, Pneumonia, and Pneumothorax Cheonan Soonchunhyang University Hospital including Tuberculosis were used. To improve performance, preprocessing was performed Center Crop while maintaining aspect ratio 1:1. Noisy Student EfficientNet B7, fine-tuning using weights learned from ImageNet, features each layer maximally utilized Multi GAP structure. result experiment, Benchmarks measured NIH showed highest performance among tested models an accuracy 85.32%, four-class predictions in had average 96.1%, sensitivity 92.2%, specificity 97.4%, inference time 0.2 s.