Optimized ResNet-18 Architecture for Multi-Class Oral Diseases Classification

In recent years, the classification of oral diseases has gained significant attention due to its influence on public health and the necessity for early and accurate diagnosis. Traditional diagnosis depends on manual clinical assessment, which can be slow and subjective. An optimized and subsequently quantized model is required to provide a faster and more consistent diagnostic support tool. This paper proposes an optimized ResNet-18 architecture for the classification of six oral diseases. The optimization process is based on removing the Rectified Linear Unit (ReLU), Batch Normalization (BN), and convolutional layers from the base ResNet-18 blocks that contain 128 filters. This reduction decreases computational complexity while maintaining feature-learning capability. The optimized model achieves a classification accuracy of 90.94% on a diverse oral disease dataset, and it surpasses several baseline deep learning models. Additionally, post-training quantization is applied to the optimized ResNet-18, reducing memory requirements by about 75% with no loss in accuracy and making it suitable for real-time use in clinical environments, especially on devices with limited computational power.