Classification of Throat Disease Using CNNs: EfficientNetB0 and ResNet50

Abstract

Throat diseases are one of the global health issues. Early diagnosis could be an effective solution to prevent more severe throat disease. Automatic diagnosis based on medical images is possible to obtain by using Convolutional Neural Networks (CNN). This study employs two pretrained models namely ResNet50 and EfficientNetB0. The dataset contained 79 throat images divided to seven classes (normal, chronic laryngitis, acute pharyngitis, chronic pharyngitis, acute tonsillitis, chronic tonsillitis, and acute tonsillopharyngitis). The study was conducted in several scenarios and implemented gradually. First scenario, seven classes were merged into four classes (normal, pharyngitis, tonsillitis, and acute tonsillopharyngitis). Second scenario, four classes were combined into three classes (normal, pharyngitis, and tonsillitis). Third scenario, three classes were grouped into two classes (normal and illness). The results indicated that both the ResNet50 and EfficientNetB0 architectures achieved the highest performance in the third scenario (two classes). Both models showed identical evaluation matrics with accuracy of 91,67%, precision of 90%, recall of 100%, and F1-score of 94,74%. Furthermore, this study suggests that a dataset with numerous classes and limited data can be addressed by merging classes, thereby increasing the data size within each class.

Key words: Classification, Throat Disease, CNN, ResNet50, EfficientNetB0.