Dento maxillo facial radiology最新文献

Objectives: This study aimed to conduct a comprehensive evaluation of general-purpose multimodal large language model (LLM) chatbots in oral radiology.

Methods: Ninety text- and image-based oral radiology questions from a Korean dental university were extracted and categorized into six educational contents and two question types. ChatGPT-4o and Gemini 2.0 Flash were evaluated with following items: accuracy with group differences across six contents (using Fisher's exact test with Bonferroni correction, p < 0.0167), answer consistency across ten repeated outputs (evaluated as the mean agreement and Fleiss' kappa coefficient), and hallucination (evaluated as the mean of the 5-point Global Quality Score assigned by two oral radiologists).

Results: Multimodal AI chatbots (ChatGPT-4o and Gemini 2.0 Flash) achieved excellent performance on text-based questions with over 80% accuracy but showed limited performance on image-based tasks, with accuracy under 30%. Additionally, image-based tasks exhibited high response variability, and hallucinations were frequently observed, providing incorrect information. These findings suggest that AI chatbots are not yet suitable for reliable use in oral radiology.

Conclusions: This study provided timely insights into the capabilities and limitations of general-purpose multimodal LLM chatbots in the oral radiology, and will serve as a foundation for more safe and effective applications of AI chatbots in the oral radiology field in the future.

Advances in knowledge: This is the first study to comprehensively assess multimodal LLM chatbots in oral radiology. It provides key insights into the performance benchmarks for AI chatbots in oral radiology, promoting the responsible and transparent integration of AI into dental education.

Objectives: The objective was to assess the accuracy of different AI tools in providing the right responses for multiple choice questions (MCQs) and the time taken to complete the responses.

Methods: The study included 80 MCQs, each with four options and a correct answer related to oral radiology used to assess the knowledge and skill across 5 different domains. The accuracy levels of ChatGPT, ChatGPT-4o (4o), Microsoft Co-pilot, DeepSeek, Gemini, and Meta AI were assessed and compared using the Chi-Square test. In addition, One-way ANOVA was used to compare the response time between different AI chatbots.

Results: Overall, Microsoft Co-pilot and ChatGPT-4o had the highest accuracy. ChatGPT had the fastest response time. Microsoft Co-pilot and DeepSeek, though not significant, had the highest accuracy for knowledge-based and skill-based queries. For accuracy on 5 domains, Microsoft Co-pilot and ChatGPT-4o had an accuracy of 100% for radiographic safety and DeepSeek was more accurate for radiographic diagnosis. Students took more time to respond than the collective time taken by AI chatbots.

Conclusion: Microsoft Co-pilot had an overall higher accuracy, responded more accurately for knowledge-based questions, and was 100% accurate for queries related to radiographic safety. ChatGPT-4o had the second highest accuracy and DeepSeek served better for radiographic diagnosis.

Advances in knowledge: This study is the first to systematically compare the accuracy and response time of multiple AI-driven tools in answering domain-specific MCQs in oral radiology. It highlights significant variability in performance across platforms, offering novel insights into the suitability of AI chatbots for educational use in dentistry.

Objectives: To evaluate the diagnostic accuracy and generalizability of artificial-intelligence (AI) models for radiographic classification of jawbone cysts and tumours, and to propose a Clinical Interpretation Score (CIS) that rates the transparency and real-world readiness of published AI tools.

Methods: Eligible studies reporting sensitivity and specificity of AI classifiers on panoramic radiographs or cone-beam CT were retrieved. Two reviewers applied JBI risk-of-bias criteria and extracted 2 x 2 tables and relevant metrics. Pooled estimates were calculated with random-effects meta-analysis; heterogeneity was quantified with I2.

Results: Nineteen studies were included, predominantly reporting convolutional neural networks. Pooled specificity was consistently high (≥0.90) across lesions, whereas sensitivity ranged widely (0.50-1.00). Stafne bone cavities achieved near-perfect metrics; ameloblastoma and odontogenic keratocyst showed moderate sensitivity (0.62-0.85) but retained high specificity. Cone-beam CT improved sensitivity relative to panoramic imaging. Substantial heterogeneity (I2 > 50% in most comparisons) reflected variable prevalence, imaging protocols and validation strategies.

Conclusions: AI models demonstrate promising diagnostic performance in classifying several jawbone lesions, though their accuracy is influenced by imaging modality, lesion type, and prevalence. Despite encouraging technical results, many studies lack transparent reporting and external validation, limiting their clinical interpretability. The Clinical Interpretation Score (CIS) provides a structured framework to evaluate the methodological transparency and clinical readiness of AI tools, helping to distinguish between technically sound models and those suitable for integration into diagnostic workflows.

This technical report presents a novel CBCT-Based Grading System for Oropharyngeal Airway Narrowing, designed to provide clinicians with a standardised, objective method to assess oropharyngeal airway narrowing using Cone-Beam Computed Tomography (CBCT). The grading system is developed based on the least surface area on axial section measurements/minimal cross-sectional area (MCA) on CBCT. It classifies oropharyngeal narrowing into five distinct grades (Grade 0 to 4). Each grade also has subcategories that correspond to specific anatomical regions-distal to the soft palate (P), distal to the base of the tongue (T), or distal to both the soft palate and the tongue (B) - and includes precise surface area ranges, contributing to better understanding. Traditional methods have commonly relied upon lateral cephalometry or supine CT; however, CBCT offers 3D mapping in a natural upright position, ensuring functional relevance of the airway assessment. Owing to its high spatial resolution, adequate contrast between the soft tissue and empty space, relatively low radiation dose compared to multidetector row CT and visibility of the upper airway by utilising a large field of view (FOV) protocol, makes CBCT a useful diagnostic tool for evaluation of the airway. The fact that CBCT is taken in a sitting or standing position, where the head is in equilibrium and orofacial, neck musculature is in voluntary control, vis-à-vis the supine position, where this control is taken over by the autonomic nervous system, and the distal part of the soft palate compresses the already narrowed airway further adds to its usefulness. CBCT imaging, with its three-dimensional mapping capabilities, allows for precise visualisation of the airway from the level of the posterior nasal spine, where the hard palate ends, extending to the epiglottis-thus measuring the oropharyngeal airway. The system is particularly useful for early detection and evaluation of conditions such as obstructive sleep apnoea, hypertrophy of the nasopharyngeal tonsils (adenoids), predicting difficult airways for ease of intubation, guiding orthognathic surgical interventions, craniofacial anomalies, and complex orthognathic surgical planning. It holds promise for integration into AI-enabled diagnostic platforms and digital imaging software, offering consistency in research and practice. This report details the rationale, grading criteria, anatomical references, and potential applications of this classification. The system offers a streamlined approach for identifying airway compromise, ultimately aiding multidisciplinary use in optimising patient outcomes.

Objectives: Panoramic radiographs (PANs) are a commonly used dental imaging modality and are complex to evaluate. This study aimed to use mouse-tracking technology to identify the visualization characteristics of experienced Oral & Maxillofacial Radiology (OMR) faculty when reviewing PANs and to assess relationships with detection accuracy.

Methods: Seventeen OMR faculty in U.S. dental schools with 5 to 38 years of experience were recruited. Participants were shown 17 PANs (5 "Normal" and 12 "Pathology") over Zoom and instructed to move their mouse cursor in sync with their eye movements. Data collection was based on video recordings with X, Y coordinates auto-detected using mouse-tracking algorithms. Parameters collected included detection accuracy, time taken, completeness of search coverage, revisits and search patterns used.

Results: Most common search patterns were Dental to Periphery (41.5%) and Periphery to Dental (37.7%). The mean accuracy score for all cases with pathology was 84.6%. Except for 2 cases with subtle findings (condylar fracture and fibrous dysplasia), mean accuracy scores ranged from 76.5% to 100%. There were no associations between search patterns and detection accuracy. Participants took longer durations and performed more complete searches with more revisits on "Normal" cases. They were more likely to use shorter search times with smaller coverage and perform single searches when correctly detecting lesions.

Conclusions: Experienced OMRs were able to detect lesions without scanning the entire PAN and within a single search, regardless of the search pattern used.

Advances in knowledge: This is the first study using mouse-tracking algorithms to evaluate visualization patterns on PANs. Experienced OMRs can conduct efficient and accurate reviews of PANs regardless of search strategy.

Objectives: This study proposes a deep convolutional neural network model that integrates B-mode and D-mode ultrasound images to classify metastatic lymph nodes in patients with oral squamous cell carcinoma.

Methods: A shared backbone network incorporating a cross-attention mechanism was employed to enhance feature-level interactions between dual-input ultrasound images. A total of six convolutional neural network architectures (VGG16, SqueezeNet, ResNet50, EfficientNet B3, ConvNext, DenseNet121) were implemented within a shared backbone framework to investigate optimal performance. For each network, diagnostic performance was compared between dual-input and single-input ultrasound. In addition, model performance was evaluated against human observers with different levels of experience.

Results: The model using DenseNet121 as a shared backbone with an integrated cross-attention layer (LNM-Net) achieved the highest classification accuracy (85.3%) when utilizing dual-input images, surpassing the diagnostic performance of residents. The cross-attention module improved feature fusion, reducing false positives by suppressing modality-specific noise.

Conclusion: LNM-Net demonstrates strong potential as a clinical decision-support tool for preoperative lymph node metastasis assessment in oral squamous cell carcinoma. Despite current limitations such as dataset size and cross-institutional variability, the model offers a promising supplementary aid, particularly in settings with limited radiological expertise.

Advances in knowledge: This study develops a novel cross-attentive network using dual-input B- and D-mode ultrasound images to classify metastatic lymph nodes in oral squamous cell carcinoma.

Objective: This research aimed to evaluate the results of cephalometric analyses obtained by artificial intelligence (AI) from the RadioCef, EasyCeph, and WebCeph platforms, and their variability due to modifications made by the user.

Methods: In this cross-sectional observational study, seventy cephalometric radiographs were analysed using the AI of the platforms. Subsequently, four examiners with different areas of expertise and levels of experience examined each landmark, correcting its location if necessary.

Results: The Pog, L1 tip, B, and Go landmarks on the RadioCef; Pn, Me, Pog, U1 tip, and UL on the EasyCeph; and Pog, Me, and B on the WebCeph showed a modification equal to or greater than 90%. More experienced examiners modified a greater number of landmarks. The repeated measures ANOVA test reported statistically significant differences concerning the SNA, SNB, ANB, SN-GoGn, FMIA, FMA, and IMPA angles (P < .05) for fully automated and semi-automated analyses. ICC values reported intra-observer agreement levels from poor (ICC = 0.27) to perfect (ICC = 1), and inter-observer agreement showed good to excellent reliability (ICC = 0.88-0.99).

Conclusions: Fully automated cephalometric analysis presents variations according to modifications made by the examiners. This represents a challenge to the knowledge of the orthodontist, influencing the diagnosis and treatment planning. Therefore, the use of augmented intelligence in cephalometric analysis is still suggested based on the results obtained for each platform.

Advances in knowledge: Cephalometric AI platforms show variability in landmark location accuracy. User modifications significantly impact automated analysis results.

Objectives: Cone beam computed tomography (CBCT) has revolutionized dental imaging due to its high spatial resolution and ability to provide detailed 3-dimensional reconstructions of dental structures. This study introduces an innovative CBCT image processing method using an oriented object detection approach integrated with a Region of Interest (RoI) transformer.

Methods: This study addresses the challenge of accurate tooth detection and classification in PAN derived from CBCT, introducing an innovative oriented object detection approach, which has not been previously applied in dental imaging. This method better aligns with the natural growth patterns of teeth, allowing for more accurate detection and classification of molars, premolars, canines, and incisors. By integrating RoI transformer, the model demonstrates relatively acceptable performance metrics compared to conventional horizontal detection methods while also offering enhanced visualization capabilities. Furthermore, post-processing techniques, including distance and greyscale value constraints, are employed to correct classification errors and reduce false positives, especially in areas with missing teeth.

Results: The experimental results indicate that the proposed method achieves an accuracy of 98.48%, a recall of 97.21%, an F1 score of 97.21%, and an mean average precision (mAP) of 98.12% in tooth detection.

Conclusions: The proposed method enhances the accuracy of tooth detection in CBCT-derived PAN by reducing background interference and improving the visualization of tooth orientation.

Objectives: This survey was conducted to identify dental radiography practices and knowledge gaps among dentists in Singapore with respect to both 2D and 3D imaging.

Methods: This was a cross-sectional survey conducted via an electronic platform from May to June 2023. All Singapore Dental Council-registered dentists were eligible to participate. We gathered data on demographics, intraoral (IO) radiography usage, cone-beam CT (CBCT) training, current CBCT usage, and interpretation.

Results: Three-hundred and five out of 2605 registered dentists completed the online survey (11.7% response rate). Radiograph positioning holders, digital imaging, and round collimation were used by most respondents for IO radiography (85.6%, 88.9%, and 89.2%, respectively). One-hundred and forty-two participants (46.6%) underwent CBCT training, and most lacked training in image acquisition. Dentists expressed interest in CBCT interpretation and the use of viewing software; 219 dentists (71.8%) are CBCT users. Most (74.9%) took 0-5 CBCT scans monthly. Implant planning was the most common indication (24.8%). Among CBCT users, 85.9% report some or all their own scans while 70.7% would engage a reporting service if available.

Conclusions: This survey provides insights in dental imaging that require increased educational efforts in Singapore. For IO radiography, there should be greater emphasis on rectangular collimation. With CBCT, there is potential for training in image acquisition, interpretation, and the use of viewing software.

Advances in knowledge: Identifying knowledge and training gaps in dental imaging is critical to ensure the safe use of ionizing radiation on patients.

Objectives: This study aimed to systematically review the current performance of large language models (LLMs) in dento-maxillofacial radiology (DMFR).

Methods: Five electronic databases were used to identify studies that developed, fine-tuned, or evaluated LLMs for DMFR-related tasks. Data extracted included study purpose, LLM type, images/text source, applied language, dataset characteristics, input and output, performance outcomes, evaluation methods, and reference standards. Customized assessment criteria adapted from the TRIPOD-LLM reporting guideline were used to evaluate the risk-of-bias in the included studies specifically regarding the clarity of dataset origin, the robustness of performance evaluation methods, and the validity of the reference standards.

Results: The initial search yielded 1621 titles, and 19 studies were included. These studies investigated the use of LLMs for tasks including the production and answering of DMFR-related qualification exams and educational questions (n = 8), diagnosis and treatment recommendations (n = 7), and radiology report generation and patient communication (n = 4). LLMs demonstrated varied performance in diagnosing dental conditions, with accuracy ranging from 37% to 92.5% and expert ratings for differential diagnosis and treatment planning between 3.6 and 4.7 on a 5-point scale. For DMFR-related qualification exams and board-style questions, LLMs achieved correctness rates between 33.3% and 86.1%. Automated radiology report generation showed moderate performance with accuracy ranging from 70.4% to 81.3%.

Conclusions: LLMs demonstrate promising potential in DMFR, particularly for diagnostic, educational, and report generation tasks. However, their current accuracy, completeness, and consistency remain variable. Further development, validation, and standardization are needed before LLMs can be reliably integrated as supportive tools in clinical workflows and educational settings.