A Proposed Framework for an Artificial Intelligence-based Clinical Decision Support System in Dento-Maxillofacial Radiology

Abstract

Artificial intelligence shows promise for automated image analysis in dento-maxillofacial radiology (DMFR), yet few systems reach routine clinical use. This thesis argues that the gap between benchmark performance and adoption stems from four limitations of deep-learning-only systems—interpretability, generalisability, trustworthiness, and explainability—and proposes an AI-based clinical decision support system (CDSS) addressing them through hybrid deep-learning/rule-based (HDLRB) architectures, modular agentic orchestration, and human-in-the-loop design.

Three empirical studies underpin the framework. A systematic review of how panoramic imaging errors are handled during machine-learning development revealed marked inconsistencies, exposing a generalisability gap that motivates automated quality assurance as a CDSS's first stage. Two further studies showed that HDLRB pipelines—pairing deep-learning segmentation with deterministic spatial analytics—can simulate expert reasoning for the surgical management of impacted mandibular third molars and for staging periodontal bone loss on cone-beam computed tomography (CBCT). Both reached strong agreement with expert consensus while preserving full transparency, every recommendation traceable to inspectable measurements and auditable rules.

The thesis synthesises these into a framework built on five principles—hybrid architecture, modular auditable orchestration, human-in-the-loop design, quality assurance as the first diagnostic stage, and determinism where possible—structured around a QA–anatomy–pathology pathway mirroring DMFR specialist reasoning. The PerioDetect Registered Report, a supplementary appendix, operationalises these as a fully auditable multi-agent system for periodontal CBCT assessment. This work shows HDLRB pipelines can match expert-level agreement on structured DMFR tasks without sacrificing the transparency clinicians need for trust, and extends to further specialties, modalities, and technologies.