Deep Learning-Based Throat Infection Detection: A Transfer Learning Approach with MobileNet, ResNet and DenseNet with Augmented Medical Image Analysis

Deep learning offers a promising avenue for automated throat infection detection. This study assessed the performance of MobileNet, ResNet, and DenseNet, three Convolutional Neural Network (CNN) models, in classifying healthy versus pharyngitis-affected throat images. Training and testing were conducted using a labeled dataset, with performance evaluated through accuracy, loss, F1-score, confusion matrices, and Grad-CAM visualizations. MobileNet, despite its computational efficiency, achieved moderate classification accuracy (approximately 75%) and exhibited high misclassification rates. ResNet demonstrated strong generalization with about 95% accuracy, effectively balancing feature extraction and classification confidence. DenseNet achieved the highest accuracy, ranging from 91–95%, though minor overfitting was observed in later epochs. Confusion matrix analysis confirmed DenseNet's superiority with the lowest false positives and false negatives, indicating it was the most reliable model. Grad-CAM visualizations highlighted varying feature extraction approaches: MobileNet focused on structural throat features, ResNet captured both local and global details, and DenseNet leveraged dense connections for enhanced feature detection. While DenseNet outperformed the other models, further optimization through regularization techniques and data augmentation could improve its stability. These findings underscore the importance of model selection in medical image classification and suggest deep learning's potential for effective, automated throat infection detection. Future research could explore hybrid architectures or ensemble learning to enhance diagnostic accuracy.