Comparative Evaluation of Vision-Language Models for Detecting and Localizing Dental Lesions from Intraoral Images

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Computer Vision

Background:

• Intraoral photo imaging is crucial for diagnosing dental caries.
• Vision-language models offer potential for automated analysis of dental images.
• Current models require evaluation for their efficiency in detecting various dental lesions.

Purpose of the Study:

• To evaluate the performance of Florence-2, PaLI-Gemma, and YOLOv8 in detecting and classifying carious and non-carious dental lesions.
• To compare the efficacy of these models using a standardized dataset of intraoral photographs.

Main Methods:

• A dataset of 172 annotated intraoral images was used, classifying lesions such as microcavitation, cavitated lesions, staining, calculus, and non-carious lesions.
• Models were trained on an 80:10:10 data split.
• Performance was assessed using mean average precision (mAP), mAP50-95, and class-specific precision and recall metrics.

Main Results:

• YOLOv8 achieved the highest overall performance with a mean average precision (mAP) of 37%, 42.3% precision, and 31.3% recall, including 100% precision for cavitation detection.
• PaLI-Gemma achieved a recall of 13% and 21% (precision not specified).
• Florence-2 yielded a mean average precision (mAP) of 10%, with 51% precision and 35% recall.

Conclusions:

• YOLOv8 significantly outperformed Florence-2 and PaLI-Gemma in detecting and classifying dental lesions from intraoral images.
• Despite multimodal capabilities, Florence-2 and PaLI-Gemma underperformed, suggesting a need for larger, more diverse datasets.
• Hybrid architectures may be necessary to improve the performance of vision-language models in dental imaging applications.