AI-Augmented Digital Diagnostics in Minimally Invasive Dentistry: A Case Report on Clinically Occult Approximal Caries

Abstract

Background: Approximal dental caries pose a persistent diagnostic challenge, as early lesions frequently remain clinically occult beneath intact enamel surfaces during routine visual–tactile inspection. This case report describes the clinical utility, limitations, and decision-making discrepancies within a multimodal digital diagnostic workflow combining intraoral radiography, near-infrared light transillumination (NILT), and deep-learning artificial intelligence (AI) radiographic analysis. Case Presentation: A 19-year-old female patient presented with provoked cold and sweet sensitivity localized to the upper left quadrant. Clinical inspection revealed visually intact accessible enamel surfaces (ICDAS 0 on tooth 25. To resolve the discrepancy between subjective symptoms and inconclusive initial visual findings, periapical radiography and NILT were performed. After the clinician-led treatment plan had been established, the radiograph was additionally processed using ClinicalAID as a parallel research analysis. Diagnostic Assessment and Management: Periapical radiography and NILT revealed approximal demineralization on teeth 24, 25, and 26. After the clinician-led treatment plan had been established, the radiograph was additionally processed with ClinicalAID as a parallel research analysis, yielding model confidence scores of 100%, 99%, and 98% for teeth 24, 25, and 26, respectively. Pre-operative therapeutic stratification remained entirely clinician-led: teeth 24 and 25 were selected for operative intervention based on provoked symptomatology and radiographic appearance of dentinal involvement, whereas tooth 26 was managed non-operatively with topical fluoridation and scheduled monitoring. Operative clinical reference assessment confirmed demineralization extending into the middle third of the dentin on tooth 25. Conclusions: High AI caries-detection confidence scores do not differentiate between lesion depth, cavitation status, or the necessity of operative intervention, as evidenced by divergent clinical pathways despite near-identical AI output scores on teeth 25 and 26. Deep-learning algorithms must function strictly as adjunctive decision-support tools within a human-in-the-loop framework (Intelligence Augmentation), with treatment planning governed by clinical symptoms, caries risk, and conventional diagnostic synthesis.