Smile aesthetic evaluation is a critical component of clinical treatment planning. As a key aesthetic indicator, the precise classification of the smile line is crucial for optimizing restorative and reconstructive treatment plans. However, classifying smile line requires accurate assessment of complex relationships among the lips, gingiva and teeth, and analysis by dentists involves a degree of subjectivity and chances for misdiagnosis. This study aimed to investigate smile line classification by comparing the performance of convolutional neural networks (CNNs) and large language models (LLMs) , as well as clinicians of varying expertise levels, in this task.

Based on the publicly available high-quality FFHQ facial dataset, a smile image annotation dataset comprising 1 000 samples was constructed following image preprocessing and standardized annotations of three types: high, medium and low smile line. Seven classic CNN models (VGG16, ResNet34, etc.) and five representative multimodal LLMs (Qwen series, LLaVA 1.5-7B) were employed for training, validation, and testing. Model performance was evaluated using accuracy, precision, recall, and F₁ scores, and compared against assessments made by clinicians of different seniority levels.

Among the seven commonly used CNN models, the ResNet152 model demonstrated optimal overall performance, achieving a mean classification accuracy of 83.30%, which significantly outperformed other CNN models and multimodal LLMs. Senior dentists achieved a classification accuracy of 83.00%, comparable to the performance of ResNet152. Heatmaps demonstrate similar attention regions between ResNet152 and dental practitioners.

CNN models demonstrated substantial clinical potential in smile line classification tasks, attaining expert-level performance. In contrast, large language models required further optimization for medical image fine-grained classification. This study provided experimental evidence and technical insights for developing intelligent aesthetic assessment systems in dentistry.