Abstract Objectives To compare final-answer accuracy of web-based large language models (LLMs) on Turkish and English image-based dental examination multiple-choice questions (deMCQs), and to describe explanation quality, reference quality, output volume, and an exploratory word-normalized water-use scenario. Methods Fifty-two image-based deMCQs (2012–2025) were evaluated in Turkish and English using ChatGPT, Gemini, and Claude, yielding 312 responses. Accuracy was coded as correct or incorrect. Explanation and reference quality were assessed using a 5-point Likert scale and treated as exploratory descriptive outcomes. Accuracy was analyzed using question-clustered binomial generalized estimating equations (GEE) with model, language, and model×language interaction terms. Output word counts were summarized, and a study-specific exploratory word-normalized water-use scenario was derived from a published exchange-level estimate. Results Overall accuracy was 56.7% (177/312). Accuracy was 59.0% in Turkish and 54.5% in English. GEE analysis showed no significant effects of model, language, or model×language interaction (p > 0.05). Descriptively, Gemini had the highest accuracy (61.5%), whereas Claude had the highest mean explanation (3.95±1.31) and reference quality (4.85±0.38) scores. Across 312 prompt–response exchanges, the models generated 34,534 words, corresponding to an exploratory word-normalized water-use proxy of 3.12–7.80 L. Conclusions Freely accessible web-based LLMs demonstrated moderate accuracy on image-based deMCQs. The water-use proxy was derived from published exchange-level assumptions and should not be interpreted as a direct measurement or validated model-specific estimate of water consumption. Descriptive differences between models and languages should not be interpreted as evidence of superiority or equivalence. These systems may support dental education; however, outputs require expert verification before educational or clinical use. Advances in knowledge This study evaluates three widely accessible web-based multimodal LLMs using an entirely image-based dental examination question set, accounts for repeated observations through question-clustered GEE, compares Turkish and English performance without equating non-significance with equivalence, and presents a transparent word-normalized environmental scenario with explicit assumptions and limitations.
Abstract Objectives The aim of this study was to investigate the feasibility of low-field dental MRI in the evaluation of the inferior alveolar nerve (IAN) and the acceleration of image acquisition with a deep learning (DL)-based sequence. Therefore, overall image quality (OIQ), image artefacts and delineation of the IAN were investigated. Methods In this single-centre retrospective study, we enrolled 39 patients who underwent dental MRI using a dedicated dental coil on a 0.55 T MRI system. A conventional T1-weighted gradient-echo 3D sequence and a T1-weighted gradient-echo 3D sequence with advanced DL-based reconstruction were acquired. Both sequences were evaluated for OIQ, image artefacts in and conspicuity of the IAN using a Likert scale from 1 to 5. Interrater agreement was assessed using weighted Cohen’s κ coefficients. As quantitative measurements, the apparent signal-to-noise ratio (aSNR) and the diameter of the IAN were determined. Results Thirty-three examinations were included, OIQ was rated higher for the accelerated sequence compared with the conventional sequence for both readers (p < 0.01). Image artefacts were lower in the accelerated sequence for both readers (p < 0.001). Accordingly, the delineation of the IAN was improved for both sides (p < 0.001) in the accelerated sequence. Interrater agreement between readers was substantial/good to almost perfect/excellent across all measures. aSNR was increased in the accelerated sequence (p < 0.001), whereas IAN diameter did not differ. Conclusions This study demonstrates that a DL-accelerated T1-weighted gradient-echo 3D sequence for dental-dedicated MRI is feasible for imaging of the IAN with improved image quality while reducing acquisition time by 50%. Advances in Knowledge Reduction of scan time and superior image quality could help establish dental-dedicated MRI as a radiation-free alternative for dentomaxillofacial imaging.
Objectives To compare the diagnostic performance of 3 multimodal artificial intelligence (AI) chatbots on oral and maxillofacial radiographic images and to examine how additional information, delivery mode, and user-suggested diagnoses affect accuracy.Methods Three AI chatbots (GPT-5.1, Gemini 3 Flash, Claude Opus 4.7) were tested on 90 cases comprising normal controls, osteomyelitis, and benign jaw lesions. Inputs were combinations of a panoramic image, a cropped panoramic image, an axial cone-beam CT (CBCT) image, or a text-based cue. Inputs were delivered all at once or sequentially. Accuracy was scored at category and specific-diagnosis levels using non-parametric tests with false-discovery-rate correction.Results With the panoramic image alone, accuracy for diseased cases was low (0%-60%) but rose to as high as 38%-92% in each model's best condition with added information. A cropped image was the most consistently beneficial additional visual input, whereas an axial CBCT image provided less improvement. GPT-5.1 recognized normal cases well but missed lesions, Gemini 3 was sensitive but less specific, and Claude 4.7 defaulted to benign diagnoses. Correct verbal cues increased accuracy, whereas a misleading cue caused a decline. Gemini 3 accepted a false benign suggestion in 96% of cases it had initially classified as normal. For benign lesions, specific-diagnosis accuracy was almost half of category-level accuracy.Conclusions Accuracy was strongly affected by model choice, information type, and user-suggested diagnoses, but no model reached a level that would be acceptable for clinical use.Advances in knowledge This multi-model comparison isolates the effects of information type, delivery mode, and sycophancy in oral and maxillofacial radiology.
Objectives To evaluate the sensitivity of 2 dental artificial intelligence (AI) systems for detecting dentinal caries using longitudinal, outcomes-anchored validation within an integrated healthcare system.Methods In this pilot external validation study, 2 commercially available AI systems were tested on full-mouth intraoral radiographic series, including bitewings, from 90 patients in a federal healthcare system. Ground truth was established using baseline clinical examinations and follow-up documentation 6-12 months later. Vendor-specific and pooled sensitivity for dentinal caries detection were calculated.Results Vendor A achieved a sensitivity of 48.5% (95% CI: 37.0-60.2), Vendor B achieved 57.5% (95% CI: 46.1-68.2), and pooled sensitivity was 53.0% (95% CI: 44.8-61.1).Conclusions Outcomes-anchored, longitudinal validation of dental AI systems for dentinal caries detection was feasible within a large integrated healthcare system. This approach provides clinically relevant sensitivity estimates grounded in real treatment outcomes. These findings are most directly applicable to large integrated healthcare systems with standardized longitudinal data capture and may not fully reflect the variability of private-practice dental settings.Advances in knowledge This study establishes a reproducible framework for evaluating AI diagnostic performance in real-world healthcare settings using longitudinal clinical outcomes rather than radiographic labels alone, demonstrating that independent validation with clinical follow-up is feasible within a large healthcare system.
Objectives: To evaluate the impact of exposure setting variations on contrast and spatial resolution (SR) in complementary metal oxide semiconductor (CMOS) and direct-conversion CMOS (DC-CMOS) intraoral sensors. Methods: Ten images of a digital dental quality assurance (DDQA) phantom were acquired for each assessed condition using 3 sensors (CMOS-I, CMOS-S, DC-CMOS) and 2 X-ray units (HH: handheld; WM: wall-mounted). HH was set at 60 kVp/2.5 mA, and WM was set at 60 kVp/7 mA and 70 kVp/7 mA. Each combined sensor-X-ray unit condition was exposed at 0.02, 0.04, 0.05, 0.08, 0.10, 0.16, 0.20, 0.25, 0.32, 0.40, 0.50, 0.63, 0.80, and 1.00 s. One observer selected 2 & times; 2 mm ROIs from the third and fifth steps of the DDQA step-wedge and measured their mean gray values using Fiji macro. Contrast was calculated by subtracting the mean gray value of the third step from the fifth step for each group. SR was determined by identifying the line pair with the highest value, in which 5 peaks and 4 valleys were distinguishable. Descriptive statistics were calculated for SR and 3-way ANOVA and Tukey's post hoc tests for contrast (alpha = 0.05). Results: Across all exposure parameters, CMOS-I exhibited the highest mean contrast, followed by DC-CMOS and CMOS-S (P < .001). HH yielded the highest mean contrast, which was significantly greater than WM (P < .001). CMOS-S exhibited the highest SR, and DC-CMOS the lowest, although SR remained constant for DC-CMOS across all exposure times. Conclusions: Sensor type, X-ray unit, and exposure time influenced image contrast and spatial resolution, while variations in kVp had minimal impact on these image quality parameters.
