Vision-Language Model as a 'Zero-Shot' Assistant for Evaluating Condylar Osseous Changes in Cone-beam Computed...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Oral and Maxillofacial Radiology
• Computer-Aided Diagnosis

Background:

• Interpreting cone-beam computed tomography (CBCT) for condylar osseous changes presents diagnostic challenges for general dental practitioners.
• The need for advanced, accessible tools to aid in the detection of temporomandibular joint abnormalities is growing.

Purpose of the Study:

• To evaluate the 'zero-shot' diagnostic performance of Vision-Language Models (VLMs) in identifying condylar abnormalities on CBCT images.
• To assess the utility of VLMs as AI assistants for general practitioners in detecting osseous changes in the mandibular condyle.
• To analyze the quality of AI-generated diagnostic reports using a standardized framework.

Main Methods:

• Analysis of 72 CBCT images (EHPN study) for internal validation and 70 images (MMDental dataset) for external validation, with dual-radiologist consensus as ground truth.
• Three frontier VLMs (Gemini-3, GPT-5.2, Qwen3-VL) were tested on representative sagittal slices without prior fine-tuning.
• Performance metrics included diagnostic accuracy, sensitivity, specificity, and balanced accuracy; report quality was assessed using the QAMAI framework, adhering to STARD-AI, CLAIM, and TRIPOD-LLM guidelines.

Main Results:

• Gemini-3 achieved the highest diagnostic accuracy (90.3% internal, 90.0% external validation), significantly outperforming GPT-5.2 (75.0%) and Qwen3-VL (55.6%).
• Gemini-3 also excelled in generating structured diagnostic reports, scoring higher across QAMAI dimensions for accuracy, clarity, and clinical utility.
• All evaluated VLMs showed a conservative diagnostic tendency.

Conclusions:

• Vision-Language Models, particularly Gemini-3, demonstrate effective 'zero-shot' capabilities for detecting condylar changes and generating high-quality diagnostic reports.
• These AI models show potential as preliminary, training-free screening aids in oral and maxillofacial radiology, enhancing diagnostic workflows for general practitioners.
• Further multicenter validation is recommended to confirm clinical utility, with cloud-based VLMs offering accessible screening assistance via simple web interfaces.