A Dual-Stage Vision–Language Framework for Dental Report Generation

Abstract

Panoramic dental radiographs are routinely used for diagnosis and treatment planning, yet their interpretation remains time-consuming and highly dependent on clinical expertise. Although Vision–Language Models (VLMs) have shown promising capabilities for medical report generation, their application to dental radiology is constrained by the scarcity of expert-written report datasets and the computational cost of domain-specific fine-tuning. This paper proposes a dual-stage framework for automated report generation from panoramic dental radiographs that integrates structured visual analysis with pretrained VLMs. In the first stage, dedicated deep learning models extract clinically relevant information through tooth detection with FDI numbering, dental pathology detection, and tooth segmentation. Specifically, two YOLO-based models perform tooth localization and pathology detection, while a U-Net-ResNet34 segmentation model generates tooth masks for constructing contour-based region-of-interest images. In the second stage, these findings are organized into a structured representation that guides pretrained VLMs through prompting. The vision models achieved an mAP at 0.50 of 0.95 for tooth detection and 0.60 for pathology detection, and 94.6% of the detected lesions were assigned to the correct tooth. Five open-weight VLMs were evaluated under four input configurations with an LLM-as-a-Judge, and the best-performing model, Qwen2.5-VL-3B, was further assessed by dental experts on 100 radiographs from an independent external dataset. Incorporating the structured findings increased the proportion of clinically relevant reports from 1.5% to 80.0–81.5%, achieved a pathology precision of up to 76.0%, and reduced the hallucination rate from 53.7% to 14.0–16.8%. These results show that structured visual findings enable small pretrained VLMs to generate clinically relevant dental reports without large-scale image–report datasets or VLM fine-tuning.