JADE-Plus: A Multimodal Agentic Retrieval-Augmented Generation Large Language Framework for Diagnostic Support in Jawbone Lesions: Development and Technical Validation Study

Diagnosing jawbone lesions in oral and maxillofacial radiology remains challenging due to overlapping radiological features and the need for integrated clinical reasoning. This study aimed to develop and validate JADE-Plus, a novel multimodal, agent-controlled retrieval-augmented generation (RAG) framework for diagnostic decision support in jawbone lesion assessment. JADE-Plus was implemented as a cloud-based, tablet-optimized system integrating a vision-language model (VLM) for panoramic radiograph analysis, a knowledge-grounded RAG module, and an agentic verification loop for diagnostic fusion and re-ranking. The system was evaluated using 40 representative jawbone lesion cases and compared with JADE, GPT-5.4, GPT-5.4 VLM, and ORAD. Performance was assessed using Top-1 and Top-3 accuracy, ablation and statistical analyses, intra-model stability, and response time. JADE-Plus achieved the highest diagnostic performance, with a Top-1 accuracy of 90% (36/40; 95% CI 76-97%) and a Top-3 accuracy of 100%, with no missed diagnoses. Cochran's Q test demonstrated significant differences among models for both Top-3 correctness (Q = 25.66, p < 0.001) and Top-1 correctness (Q = 27.55, p < 0.001). Post-hoc McNemar tests with the Benjamini-Hochberg correction showed that JADE-Plus significantly outperformed other models, particularly under the stricter Top-1 evaluation. Ablation analysis showed that the agentic verification stage improved Top-1 and Top-3 accuracy by 15 and 8 percentage points, respectively. JADE-Plus achieved the highest reproducibility (mean Jaccard similarity 0.97 ± 0.12) while maintaining a mean response time of 33 ± 1.5 s per case. JADE-Plus demonstrated superior diagnostic accuracy, stability, and reproducibility compared with baseline systems, supporting multimodal agentic RAG frameworks for jawbone lesion diagnosis.