<![CDATA[Dental and oral diseases represent a high-prevalence global health issue, yet their management still relies heavily on the subjective visual inspections of medical professionals. While automated diagnostic systems exist, previous studies predominantly employ conventional Convolutional Neural Networks (CNNs) that struggle to capture global anatomical dependencies, or standard Vision Transformers (ViTs) whose massive parameter counts hinder deployment on clinical edge devices. Furthermore, existing research is frequently constrained by limited disease classes and fails to explicitly resolve severe clinical data imbalance. To bridge these gaps, this study proposes a comprehensive multi-class oral disease image classification system using MobileViT, a lightweight hybrid architecture that efficiently combines local CNN convolutions with global transformer attention mechanisms. Evaluated on a large-scale dataset encompassing six disease classes calculus, dental caries, gingivitis, aphthous ulcers, tooth discoloration, and hypodontia, the inherent class imbalance is algorithmically addressed through a WeightedRandomSampler integrated with multi-level data augmentation utilizing RandAugment and RandomErasing. The dataset is partitioned into a 70:15:15 ratio for training, validation, and testing. Experimental results demonstrate that the proposed model achieves an accuracy of 93.61%, precision of 94.76%, recall of 93.61%, and an F1-score of 93.75% on the test set. An ablation study reveals that the combination of augmentation and sample weighting improves the F1-score by 4.2 points compared to the baseline without specific treatments. Furthermore, MobileViT explicitly outperforms conventional architectures including ResNet50, EfficientNetB0, and MobileNetV3. This research demonstrates that lightweight hybrid vision transformers can effectively resolve prior representational and imbalanced data limitations for clinical oral disease classification.]]>
