Detecting caries lesions of different radiographic extension on bitewings using deep learning

Machine Learning and Intelligent Diagnostics in Dental and Orofacial Pain Management: A Systematic Review

Purpose

The study explored the clinical influence, effectiveness, limitations, and human comparison outcomes of machine learning in diagnosing (1) dental diseases, (2) periodontal diseases, (3) trauma and neuralgias, (4) cysts and tumors, (5) glandular disorders, and (6) bone and temporomandibular joint as possible causes of dental and orofacial pain.

Method

Scopus, PubMed, and Web of Science (all databases) were searched by 2 reviewers until 29^th October 2020. Articles were screened and narratively synthesized according to PRISMA-DTA guidelines based on predefined eligibility criteria. Articles that made direct reference test comparisons to human clinicians were evaluated using the MI-CLAIM checklist. The risk of bias was assessed by JBI-DTA critical appraisal, and certainty of the evidence was evaluated using the GRADE approach. Information regarding the quantification method of dental pain and disease, the conditional characteristics of both training and test data cohort in the machine learning, diagnostic outcomes, and diagnostic test comparisons with clinicians, where applicable, were extracted.

Results

34 eligible articles were found for data synthesis, of which 8 articles made direct reference comparisons to human clinicians. 7 papers scored over 13 (out of the evaluated 15 points) in the MI-CLAIM approach with all papers scoring 5+ (out of 7) in JBI-DTA appraisals. GRADE approach revealed serious risks of bias and inconsistencies with most studies containing more positive cases than their true prevalence in order to facilitate machine learning. Patient-perceived symptoms and clinical history were generally found to be less reliable than radiographs or histology for training accurate machine learning models. A low agreement level between clinicians training the models was suggested to have a negative impact on the prediction accuracy. Reference comparisons found nonspecialized clinicians with less than 3 years of experience to be disadvantaged against trained models.

Conclusion

Machine learning in dental and orofacial healthcare has shown respectable results in diagnosing diseases with symptomatic pain and with improved future iterations and can be used as a diagnostic aid in the clinics. The current review did not internally analyze the machine learning models and their respective algorithms, nor consider the confounding variables and factors responsible for shaping the orofacial disorders responsible for eliciting pain.

Impact of Image Context on Deep Learning for Classification of Teeth on Radiographs

Objectives: We aimed to assess the impact of image context information on the accuracy of deep learning models for tooth classification on panoramic dental radiographs. Methods: Our dataset contained 5008 panoramic radiographs with a mean number of 25.2 teeth per image. Teeth were segmented bounding-box-wise and classified by one expert; this was validated by another expert. Tooth segments were cropped allowing for different context; the baseline size was 100% of each box and was scaled up to capture 150%, 200%, 250% and 300% to increase context. On each of the five generated datasets, ResNet-34 classification models were trained using the Adam optimizer with a learning rate of 0.001 over 25 epochs with a batch size of 16. A total of 20% of the data was used for testing; in subgroup analyses, models were tested only on specific tooth types. Feature visualization using gradient-weighted class activation mapping (Grad-CAM) was employed to visualize salient areas. Results: F1-scores increased monotonically from 0.77 in the base-case (100%) to 0.93 on the largest segments (300%; p = 0.0083; Mann-Kendall-test). Gains in accuracy were limited between 200% and 300%. This behavior was found for all tooth types except canines, where accuracy was much higher even for smaller segments and increasing context yielded only minimal gains. With increasing context salient areas were more widely distributed over each segment; at maximum segment size, the models assessed minimum 3-4 teeth as well as the interdental or inter-arch space to come to a classification. Conclusions: Context matters; classification accuracy increased significantly with increasing context.

Barriers and Enablers for Artificial Intelligence in Dental Diagnostics: A Qualitative Study

The present study aimed to identify barriers and enablers for the implementation of artificial intelligence (AI) in dental, specifically radiographic, diagnostics. Semi-structured phone interviews with dentists and patients were conducted between the end of May and the end of June 2020 (convenience/snowball sampling). A questionnaire developed along the Theoretical Domains Framework (TDF) and the Capabilities, Opportunities and Motivations influencing Behaviors model (COM-B) was used to guide interviews. Mayring's content analysis was employed to point out barriers and enablers. We identified 36 barriers, conflicting themes or enablers, covering nine of the fourteen domains of the TDF and all three determinants of behavior (COM). Both stakeholders emphasized chances and hopes for AI. A range of enablers for implementing AI in dental diagnostics were identified (e.g., the chance for higher diagnostic accuracy, a reduced workload, more comprehensive reporting and better patient-provider communication). Barriers related to reliance on AI and responsibility for medical decisions, as well as the explainability of AI and the related option to de-bug AI applications, emerged. Decision-makers and industry may want to consider these aspects to foster implementation of AI in dentistry.

Generalizability of deep learning models for dental image analysis

We assessed the generalizability of deep learning models and how to improve it. Our exemplary use-case was the detection of apical lesions on panoramic radiographs. We employed two datasets of panoramic radiographs from two centers, one in Germany (Charité, Berlin, n = 650) and one in India (KGMU, Lucknow, n = 650): First, U-Net type models were trained on images from Charité (n = 500) and assessed on test sets from Charité and KGMU (each n = 150). Second, the relevance of image characteristics was explored using pixel-value transformations, aligning the image characteristics in the datasets. Third, cross-center training effects on generalizability were evaluated by stepwise replacing Charite with KGMU images. Last, we assessed the impact of the dental status (presence of root-canal fillings or restorations). Models trained only on Charité images showed a (mean ± SD) F1-score of 54.1 ± 0.8% on Charité and 32.7 ± 0.8% on KGMU data (p < 0.001/t-test). Alignment of image data characteristics between the centers did not improve generalizability. However, by gradually increasing the fraction of KGMU images in the training set (from 0 to 100%) the F1-score on KGMU images improved (46.1 ± 0.9%) at a moderate decrease on Charité images (50.9 ± 0.9%, p < 0.01). Model performance was good on KGMU images showing root-canal fillings and/or restorations, but much lower on KGMU images without root-canal fillings and/or restorations. Our deep learning models were not generalizable across centers. Cross-center training improved generalizability. Noteworthy, the dental status, but not image characteristics were relevant. Understanding the reasons behind limits in generalizability helps to mitigate generalizability problems.

Imaging modalities to inform the detection and diagnosis of early caries

BACKGROUND

The detection and diagnosis of caries at the earliest opportunity is fundamental to the preservation of tooth tissue and maintenance of oral health. Radiographs have traditionally been used to supplement the conventional visual-tactile clinical examination. Accurate, timely detection and diagnosis of early signs of disease could afford patients the opportunity of less invasive treatment with less destruction of tooth tissue, reduce the need for treatment with aerosol-generating procedures, and potentially result in a reduced cost of care to the patient and to healthcare services.

OBJECTIVES

To determine the diagnostic accuracy of different dental imaging methods to inform the detection and diagnosis of non-cavitated enamel only coronal dental caries.

SEARCH METHODS

Cochrane Oral Health's Information Specialist undertook a search of the following databases: MEDLINE Ovid (1946 to 31 December 2018); Embase Ovid (1980 to 31 December 2018); US National Institutes of Health Ongoing Trials Register (ClinicalTrials.gov, to 31 December 2018); and the World Health Organization International Clinical Trials Registry Platform (to 31 December 2018). We studied reference lists as well as published systematic review articles.

SELECTION CRITERIA

We included diagnostic accuracy study designs that compared a dental imaging method with a reference standard (histology, excavation, enhanced visual examination), studies that evaluated the diagnostic accuracy of single index tests, and studies that directly compared two or more index tests. Studies reporting at both the patient or tooth surface level were included. In vitro and in vivo studies were eligible for inclusion. Studies that explicitly recruited participants with more advanced lesions that were obviously into dentine or frankly cavitated were excluded. We also excluded studies that artificially created carious lesions and those that used an index test during the excavation of dental caries to ascertain the optimum depth of excavation.

DATA COLLECTION AND ANALYSIS

Two review authors extracted data independently and in duplicate using a standardised data extraction form and quality assessment based on QUADAS-2 specific to the clinical context. Estimates of diagnostic accuracy were determined using the bivariate hierarchical method to produce summary points of sensitivity and specificity with 95% confidence regions. Comparative accuracy of different radiograph methods was conducted based on indirect and direct comparisons between methods. Potential sources of heterogeneity were pre-specified and explored visually and more formally through meta-regression.

MAIN RESULTS

We included 104 datasets from 77 studies reporting a total of 15,518 tooth sites or surfaces. The most frequently reported imaging methods were analogue radiographs (55 datasets from 51 studies) and digital radiographs (42 datasets from 40 studies) followed by cone beam computed tomography (CBCT) (7 datasets from 7 studies). Only 17 studies were of an in vivo study design, carried out in a clinical setting. No studies were considered to be at low risk of bias across all four domains but 16 studies were judged to have low concern for applicability across all domains. The patient selection domain had the largest number of studies judged to be at high risk of bias (43 studies); the index test, reference standard, and flow and timing domains were judged to be at high risk of bias in 30, 12, and 7 studies respectively. Studies were synthesised using a hierarchical bivariate method for meta-analysis. There was substantial variability in the results of the individual studies, with sensitivities that ranged from 0 to 0.96 and specificities from 0 to 1.00. For all imaging methods the estimated summary sensitivity and specificity point was 0.47 (95% confidence interval (CI) 0.40 to 0.53) and 0.88 (95% CI 0.84 to 0.92), respectively. In a cohort of 1000 tooth surfaces with a prevalence of enamel caries of 63%, this would result in 337 tooth surfaces being classified as disease free when enamel caries was truly present (false negatives), and 43 tooth surfaces being classified as diseased in the absence of enamel caries (false positives). Meta-regression indicated that measures of accuracy differed according to the imaging method (Chi²(4) = 32.44, P < 0.001), with the highest sensitivity observed for CBCT, and the highest specificity observed for analogue radiographs. None of the specified potential sources of heterogeneity were able to explain the variability in results. No studies included restored teeth in their sample or reported the inclusion of sealants. We rated the certainty of the evidence as low for sensitivity and specificity and downgraded two levels in total for risk of bias due to limitations in the design and conduct of the included studies, indirectness arising from the in vitro studies, and the observed inconsistency of the results.

AUTHORS' CONCLUSIONS

The design and conduct of studies to determine the diagnostic accuracy of methods to detect and diagnose caries in situ are particularly challenging. Low-certainty evidence suggests that imaging for the detection or diagnosis of early caries may have poor sensitivity but acceptable specificity, resulting in a relatively high number of false-negative results with the potential for early disease to progress. If left untreated, the opportunity to provide professional or self-care practices to arrest or reverse early caries lesions will be missed. The specificity of lesion detection is however relatively high, and one could argue that initiation of non-invasive management (such as the use of topical fluoride), is probably of low risk. CBCT showed superior sensitivity to analogue or digital radiographs but has very limited applicability to the general dental practitioner. However, given the high-radiation dose, and potential for caries-like artefacts from existing restorations, its use cannot be justified in routine caries detection. Nonetheless, if early incidental carious lesions are detected in CBCT scans taken for other purposes, these should be reported. CBCT has the potential to be used as a reference standard in diagnostic studies of this type. Despite the robust methodology applied in this comprehensive review, the results should be interpreted with some caution due to shortcomings in the design and execution of many of the included studies. Future research should evaluate the comparative accuracy of different methods, be undertaken in a clinical setting, and focus on minimising bias arising from the use of imperfect reference standards in clinical studies.

Better Reporting of Studies on Artificial Intelligence: CONSORT-AI and Beyond

An increasing number of studies on artificial intelligence (AI) are published in the dental and oral sciences. The reporting, but also further aspects of these studies, suffer from a range of limitations. Standards towards reporting, like the recently published Consolidated Standards of Reporting Trials (CONSORT)-AI extension can help to improve studies in this emerging field, and the Journal of Dental Research (JDR) encourages authors, reviewers, and readers to adhere to these standards. Notably, though, a wide range of aspects beyond reporting, located along various steps of the AI lifecycle, should be considered when conceiving, conducting, reporting, or evaluating studies on AI in dentistry.

Peri-Implant Bone Loss Measurement Using a Region-Based Convolutional Neural Network on Dental Periapical Radiographs

Determining the peri-implant marginal bone level on radiographs is challenging because the boundaries of the bones around implants are often unclear or the heights of the buccal and lingual bone levels are different. Therefore, a deep convolutional neural network (CNN) was evaluated for detecting the marginal bone level, top, and apex of implants on dental periapical radiographs. An automated assistant system was proposed for calculating the bone loss percentage and classifying the bone resorption severity. A modified region-based CNN (R-CNN) was trained using transfer learning based on Microsoft Common Objects in Context dataset. Overall, 708 periapical radiographic images were divided into training (n = 508), validation (n = 100), and test (n = 100) datasets. The training dataset was randomly enriched by data augmentation. For evaluation, average precision, average recall, and mean object keypoint similarity (OKS) were calculated, and the mean OKS values of the model and a dental clinician were compared. Using detected keypoints, radiographic bone loss was measured and classified. No statistically significant difference was found between the modified R-CNN model and dental clinician for detecting landmarks around dental implants. The modified R-CNN model can be utilized to measure the radiographic peri-implant bone loss ratio to assess the severity of peri-implantitis.

Evaluation of the re-mineralization capacity of a gold nanoparticle-based dental varnish: An in vitro study

Background

Dental caries is an infectious microbial disease caused by acidogenic bacteria. It leads to the dissolution of enamel, dentin, and cementum. Enamel demineralization is often appreciated as' 'White Chalky lesions or Chalky enamel''. Standard procedures for protection of these teeth are fissure sealing and topical fluoride application. A varnish is generally a material in which a resin such as copal is dissolved within an organic solvent such as ethanol. Gold is one of the most biocompatible dental materials. Gold nanoparticles were biosynthesised using aspartic acid in previous studies.

Aim and Objectives

To prepare a gold nanoparticle based dental varnish and to evaluate its re-mineralizing capacity.

Materials and Methods

Gold nanoparticle dental varnish was prepared using all the necessary constituents. This newly prepared dental varnish was compared with G.C Fuji/SnF2 dental varnish. Demineralizing capacity of the dental varnishes were analysed. The tooth specimens were prepared according to methodology and mounted on resin blocks. They were subjected to demineralization remineralization cycles. ICP-OES and Knoop Hardness tests were performed.

Results

AuNP dental varnish had a satisfactory remineralization effect on demineralised enamel. For calcium analysis, the AuNP group showed significantly more total calcium loss when compared to the SnF2 group (P < 0.05) and was statistically significant. For phosphorus analysis, AuNP group showed significantly greater net phosphorus loss compared to the SnF2 group (P < 0.05) and was statistically significant. It was also observed that the KHN increased more significantly in Group A (SnF2) as compared to Group B (AuNP) and was also statistically significant (P < 0.05).

Conclusion

AuNP dental varnish showed considerable re mineralizing property; however, it was not superior to dental varnishes like SnF2 dental varnish. Further research needs to be done in vitro to better modify the AuNP dental varnish before proceeding to in vivo studies.

Artificial intelligence in dental research: Checklist for authors, reviewers, readers

OBJECTIVES

The number of studies employing artificial intelligence (AI), specifically machine and deep learning, is growing fast. The majority of studies suffer from limitations in planning, conduct and reporting, resulting in low robustness, reproducibility and applicability. We here present a consented checklist on planning, conducting and reporting of AI studies for authors, reviewers and readers in dental research.

METHODS

Lending from existing reviews, standards and other guidance documents, an initial draft of the checklist and an explanatory document were derived and discussed among the members of IADR's e-oral network and the ITU/WHO focus group "Artificial Intelligence for Health (AI4H)". The checklist was consented by 27 group members via an e-Delphi process.

RESULTS

Thirty-one items on planning, conducting and reporting of AI studies were agreed on. These involve items on the studies' wider goal, focus, design and specific aims, data sampling and reporting, sample estimation, reference test construction, model parameters, training and evaluation, uncertainty and explainability, performance metrics and data partitions.

CONCLUSION

Authors, reviewers and readers should consider this checklist when planning, conducting, reporting and evaluating studies on AI in dentistry.

CLINICAL SIGNIFICANCE

Current studies on AI in dentistry show considerable weaknesses, hampering their replication and application. This checklist may help to overcome this issue and advance AI research as well as facilitate a debate on standards in this fields.

Detecting white spot lesions on dental photography using deep learning: A pilot study

OBJECTIVES

We aimed to apply deep learning to detect white spot lesions in dental photographs.

METHODS

Using 434 photographic images of 51 patients, a dataset of 2781 cropped tooth segments was generated. Pixelwise annotations of sound enamel as well as fluorotic, carious or other types of hypomineralized lesions were generated by experts and assessed by an independent second reviewer. The union of the reviewed annotations were used to segment the hard tissues (region-of-interest, ROI) of each image. SqueezeNet was employed for modelling. We trained models to detect (1) any white spot lesions, (2) fluorotic lesions and (3) other-than-fluorotic lesions. Modeling was performed on both the cropped and the ROI images and using ten-times repeated five-fold cross-validation. Feature visualization was applied to visualize salient areas.

RESULTS

Lesion prevalence was 37 %; the majority of lesions (24 %) were fluorotic. None of the metrics differed significantly between the models trained on cropped and ROI imagery (p > 0.05/t-test). Mean accuracies ranged between 0.81-0.84, without significant differences between models trained to detect any, fluorotic or other-than-fluorotic lesions (p > 0.05). Specificities were 0.85-0.86; sensitivities were lower (0.58-0.66). Models to detect any lesions showed positive/negative predictive values (PPV/NPV) between 0.77-0.80, those to detect fluorotic lesions 0.67 (PPV) to 0.86 (NPV), and those to detect other-than-fluorotic lesions 0.46 (PPV) to 0.93 (NPV). Light reflections were the main reason for false positive detections.

CONCLUSIONS

Deep learning showed satisfying accuracy to detect white spot lesions, particularly fluorosis. Some models showed limited stability given the small sample available.

CLINICAL SIGNIFICANCE

Deep learning is suitable for automated classification of retro- or prospectively collected imagery and may assist practitioners in discriminating white spot lesions. Future studies should expand the scope into more granular multi-class detections on a larger and more generalizable dataset.

Promises and perils of artificial intelligence in dentistry

Artificial intelligence (AI) is a subdiscipline of computer science that has made substantial progress in medicine and there is a growing body of AI research in dentistry. Dentists should have an understanding of the foundational concepts and the ability to critically evaluate dental research in AI. Machine learning (ML) is a subfield of AI that most dental AI research is dedicated to. The most prolific area of ML research is automated interpretation of dental imaging. Other areas include providing treatment recommendations, predicting future disease and treatment outcomes. The research impact is limited by small datasets that do not harness the positive correlation between very large datasets and ML performance. There is also a need to standardize research methodologies and utilize performance metrics that are appropriate for the clinical context. In addition to research challenges, this article discusses the ethical, legal and logistical considerations associated with implementation in clinical practice. This includes explainable AI, model bias, data privacy and security. The future implications of AI in dentistry involve a promise for a novel form of practicing dentistry however, the effect of AI on patient outcomes is yet to be determined.

Cost-effectiveness of Artificial Intelligence for Proximal Caries Detection

Artificial intelligence (AI) can assist dentists in image assessment, for example, caries detection. The wider health and cost impact of employing AI for dental diagnostics has not yet been evaluated. We compared the cost-effectiveness of proximal caries detection on bitewing radiographs with versus without AI. U-Net, a fully convolutional neural network, had been trained, validated, and tested on 3,293, 252, and 141 bitewing radiographs, respectively, on which 4 experienced dentists had marked carious lesions (reference test). Lesions were stratified for initial lesions (E1/E2/D1, presumed noncavitated, receiving caries infiltration if detected) and advanced lesions (D2/D3, presumed cavitated, receiving restorative care if detected). A Markov model was used to simulate the consequences of true- and false-positive and true- and false-negative detections, as well as the subsequent decisions over the lifetime of patients. A German mixed-payers perspective was adopted. Our health outcome was tooth retention years. Costs were measured in 2020 euro. Monte-Carlo microsimulations and univariate and probabilistic sensitivity analyses were conducted. The incremental cost-effectiveness ratio (ICER) and the cost-effectiveness acceptability at different willingness-to-pay thresholds were quantified. AI showed an accuracy of 0.80; dentists’ mean accuracy was significantly lower at 0.71 (minimum–maximum: 0.61–0.78, P < 0.05). AI was significantly more sensitive than dentists (0.75 vs. 0.36 [0.19–0.65]; P = 0.006), while its specificity was not significantly lower (0.83 vs. 0.91 [0.69–0.98]; P > 0.05). In the base-case scenario, AI was more effective (tooth retention for a mean 64 [2.5%–97.5%: 61–65] y) and less costly (298 [244–367] euro) than assessment without AI (62 [59–64] y; 322 [257–394] euro). The ICER was −13.9 euro/y (i.e., AI saved money at higher effectiveness). In the majority (>77%) of all cases, AI was less costly and more effective. Applying AI for caries detection is likely to be cost-effective, mainly as fewer lesions remain undetected. Notably, this cost-effectiveness requires dentists to manage detected early lesions nonrestoratively.