Hierarchical Reasoning Gates for Multimodal Large Language Models in Dental Image Interpretation: A Prompt-Level Protocol for Improving Accuracy.
Authors
Affiliations (6)
Affiliations (6)
- School of Health Science, State University of Amazonas, Manaus 69065-001, Brazil.
- Service de Chirurgie Orthopédique et Traumatologique, Centre Hospitalier Universitaire d'Orléans, 45100 Orléans, France.
- Innovative Therapies and Nanomedicine (InTheRNA), Inserm U1364, University of Orléans, 45100 Orléans, France.
- Faculty of Medicine, Lazarski University, 02-662 Warszawa, Poland.
- Université Catholique de Louvain, 1348 Louvain-La-Neuve, Belgium.
- Department of Surgical Dentistry and Maxillofacial Surgery, Danylo Halytsky Lviv National Medical University, 79010 Lviv, Ukraine.
Abstract
Multimodal large language models (MLLMs) have been used to support dental image interpretation, but performance may depend on prompt structure. We proposed Hierarchical Reasoning Gates (HRG), a prompt-level protocol organizing image interpretation into sequential stages with validation before progression. Forty anonymized panoramic radiographs of impacted mandibular molars were analyzed using three prompt architectures: P1, zero-shot natural language; P2, a five-step structured workflow followed by a final internal-validation checkpoint; and P3, HRG with 10 macro-stages and 35 procedural control blocks. Outputs were compared with expert radiology reference reports using five predefined variables. Mean Total Score was 3.52 ± 0.85, 3.58 ± 0.71, and 4.35 ± 0.74 for P1, P2, and P3, respectively; Full Accuracy was 10.0%, 5.0%, and 50.0%. Reports with errors in three or more variables occurred in 12.5%, 7.5%, and 0%, respectively. Only Apparent Anatomical Relationship differed significantly across prompts after Holm adjustment. Test-retest repeatability was high for several variables but lower for Apparent Anatomical Relationship in P3. HRG was associated with higher image-interpretation performance and fewer complex error patterns, although higher performance did not uniformly translate into greater repeatability. These findings support gated prompt design for controlling dependent steps in MLLM-based image analysis, while clinical use still requires expert supervision.