Machine learning-based medical imaging diagnosis in patients with temporomandibular disorders: a diagnostic test accuracy systematic review and meta-analysis

OBJECTIVES

Temporomandibular disorders (TMDs) are the second most common musculoskeletal condition which are challenging tasks for most clinicians. Recent research used machine learning (ML) algorithms to diagnose TMDs intelligently. This study aimed to systematically evaluate the quality of these studies and assess the diagnostic accuracy of existing models.

MATERIALS AND METHODS

Twelve databases (Europe PMC, Embase, etc.) and two registers were searched for published and unpublished studies using ML algorithms on medical images. Two reviewers extracted the characteristics of studies and assessed the methodological quality using the QUADAS-2 tool independently.

RESULTS

A total of 28 studies (29 reports) were included: one was at unclear risk of bias and the others were at high risk. Thus the certainty of evidence was quite low. These studies used many types of algorithms including 8 machine learning models (logistic regression, support vector machine, random forest, etc.) and 15 deep learning models (Resnet152, Yolo v5, Inception V3, etc.). The diagnostic accuracy of a few models was relatively satisfactory. The pooled sensitivity and specificity were 0.745 (0.660-0.814) and 0.770 (0.700-0.828) in random forest, 0.765 (0.686-0.829) and 0.766 (0.688-0.830) in XGBoost, and 0.781 (0.704-0.843) and 0.781 (0.704-0.843) in LightGBM.

CONCLUSIONS

Most studies had high risks of bias in Patient Selection and Index Test. Some algorithms are relatively satisfactory and might be promising in intelligent diagnosis. Overall, more high-quality studies and more types of algorithms should be conducted in the future.

CLINICAL RELEVANCE

We evaluated the diagnostic accuracy of the existing models and provided clinicians with much advice about the selection of algorithms. This study stated the promising orientation of future research, and we believe it will promote the intelligent diagnosis of TMDs.

Influence of alveolar ridge morphology and guide-hole design on the accuracy of static Computer-Assisted Implant Surgery with two implant macro-designs: An in vitro study

OBJECTIVES

The primary aim of this in vitro study was to evaluate the influence of alveolar ridge morphologies on the accuracy of static Computer-Assisted Implant Surgery (sCAIS). The secondary aims were to evaluate the influence of guide-hole design and implant macro-design on the accuracy of the final implant position.

METHODS

Eighteen standardized partially edentulous maxillary models with two different types of alveolar ridge morphologies were used. Each model was scanned via cone beam computer tomography prior to implant placement and scanned with a laboratory scanner prior to and following implant placement using sCAIS. The postsurgical scans were superimposed on the initial treatment planning position to measure the deviations between planned and postsurgical implant positions.

RESULTS

Seventy-two implants were equally distributed to the study groups. Implants placed in healed alveolar ridges showed significantly lower mean deviations at the crest (0.36 ± 0.17 mm), apex (0.69 ± 0.36 mm), and angular deviation (1.86 ± 0.99°), compared to implants placed in fresh extraction sites (0.80 ± 0.29 mm, 1.61 ± 0.59 mm, and 4.33 ± 1.87°; all p<0.0001). Implants placed with a sleeveless guide-hole design demonstrated significantly lower apical (1.02 ± 0.66 mm) and angular (2.72 ± 1.93°) deviations compared to those placed with manufacturer's sleeves (1.27 ± 0.67 mm; p = 0.01, and 3.46 ± 1.9°; p = 0.02). Deep-threaded tapered bone level implants exhibited significantly lower deviations at the crest (0.49 ± 0.28 mm), apex (0.97 ± 0.63 mm), and angular deviations (2.63 ± 1.85°) compared to shallow-threaded parallel-walled bone level implants (0.67 ± 0.34 mm; p = 0.0005, 1.32 ± 0.67 mm; p = 0.003, and 3.56 ± 1.93°; p = 0.01).

CONCLUSIONS

The accuracy of the final implant position with sCAIS is determined by the morphology of the alveolar ridge, the design of the guide holes, and the macrodesign of the implant.

CLINICAL SIGNIFICANCE

Higher accuracy in the final implant position was observed with implants placed in healed alveolar ridge morphologies, in implants with deep-threaded tapered macro-design, and when sleeveless surgical guide holes were used.

A generative adversarial inpainting network to enhance prediction of periodontal clinical attachment level

OBJECTIVES

Bone level as measured by clinical attachment levels (CAL) are critical findings that determine the diagnosis of periodontal disease. Deep learning algorithms are being used to determine CAL which aid in the diagnosis of periodontal disease. However, the limited field-of-view of bitewing x-rays poses a challenge for convolutional neural networks (CNN) because out-of-view anatomy cannot be directly considered. This study presents an inpainting algorithm using generative adversarial networks (GANs) coupled with partial convolutions to predict out-of-view anatomy to enhance CAL prediction accuracy.

METHODS

Retrospective purposive sampling of cases with healthy periodontium and diseased periodontium with bitewing and periapical radiographs and clinician recorded CAL were utilized. Data utilized was from July 1, 2016 through January 30, 2020. 80,326 images were used for training, 12,901 images were used for validation and 10,687 images were used to compare non-inpainted methods to inpainted methods for CAL predictions. Statistical analyses were mean bias error (MBE), mean absolute error (MAE) and Dunn's pairwise test comparing CAL at p=0.05.

RESULTS

Comparator p-values demonstrated statistically significant improvement in CAL prediction accuracy between corresponding inpainted and non-inpainted methods with MAE of 1.04 mm and 1.50 mm respectively. The Dunn's pairwise test indicated statistically significant improvement in CAL prediction accuracy between inpainted methods compared to their non-inpainted counterparts, with the best performing methods achieving a Dunn's pairwise value of -63.89.

CONCLUSIONS

This study demonstrates the superiority of using a generative adversarial inpainting network with partial convolutions to predict CAL from bitewing and periapical images.

CLINICAL SIGNIFICANCE

Artificial intelligence was developed and utilized to predict clinical attachment level compared to clinical measurements. A generative adversarial inpainting network with partial convolutions was developed, tested and validated to predict clinical attachment level. The inpainting approach was found to be superior to non-inpainted methods and within the 1mm clinician-determined measurement standard.

Automated chart filing on panoramic radiographs using deep learning

OBJECTIVE

The aim of this study is to automatically detect, segment and label teeth, crowns, fillings, root canal fillings, implants and root remnants on panoramic radiographs (PR(s)).

MATERIAL AND METHODS

As a reference, 2000 PR(s) were manually annotated and labeled. A deep-learning approach based on mask R-CNN with Resnet-50 in combination with a rule-based heuristic algorithm and a combinatorial search algorithm was trained and validated on 1800 PR(s). Subsquently, the trained algorithm was applied onto a test set consisting of 200 PR(s). F1 scores, as a measure of accuracy, were calculated to quantify the degree of similarity between the annotated ground-truth and the model predictions. The F1-score considers the harmonic mean of precison (positive predictive value) and recall (specificity).

RESULTS

The proposes method achieved F1 scores up to 0.993, 0.952 and 0.97 for detection, segmentation and labeling, respectivley.

CONCLUSION

The proposed method forms a promising foundation for the further development of automatic chart filing on PR(s).

CLINICAL SIGNIFICANCE

Deep learning may assist clinicians in summarizing the radiological findings on panoramic radiographs. The impact of using such models in clinical practice should be explored.

Standardization of three-dimensional pose of cylindrical implants from intraoral radiographs: a preliminary study

Background

To introduce a theoretical solution to a posteriori describe the pose of a cylindrical dental fixture as appearing on radiographs; to experimentally validate the method described.

Methods

The pose of a conventional dental implant was described by a triplet of angles (phi-pitch, theta-roll, and psi-yaw) which was calculated throughout vector analysis. Radiographic- and simulated-image obtained with an algorithm were compared to test effectiveness, reproducibility, and accuracy of the method. The length of the dental implant as appearing on the simulated image was calculated by the trigonometric function and then compared with real length as it appeared on a two-dimensional radiograph.

Results

Twenty radiographs were analyzed for the present in silico and retrospective study. Among 40 fittings, 37 resulted as resolved with residuals ≤ 1 mm. Similar results were obtained for radiographic and simulated implants with absolute errors of − 1.1° ± 3.9° for phi; − 0.9° ± 4.1° for theta; 0° ± 1.1° for psi. The real and simulated length of the implants appeared to be heavily correlated. Linear dependence was verified by the results of the robust linear regression: 0.9757 (slope), + 0.1344 mm (intercept), and an adjusted coefficient of determination of 0.9054.

Conclusions

The method allowed clinicians to calculate, a posteriori, a single real triplet of angles (phi, theta, psi) by analyzing a two-dimensional radiograph and to identify cases where standardization of repeated intraoral radiographies was not achieved. The a posteriori standardization of two-dimensional radiographs could allowed the clinicians to minimize the patient’s exposure to ionizing radiations for the measurement of marginal bone levels around dental implants.

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.

Fusion of intra-oral scans in cone-beam computed tomography scans

PURPOSE

The purpose of this study was to evaluate the clinical accuracy of the fusion of intra-oral scans in cone-beam computed tomography (CBCT) scans using two commercially available software packages.

MATERIALS AND METHODS

Ten dry human skulls were subjected to structured light scanning, CBCT scanning, and intra-oral scanning. Two commercially available software packages were used to perform fusion of the intra-oral scans in the CBCT scan to create an accurate virtual head model: IPS CaseDesigner® and OrthoAnalyzer™. The structured light scanner was used as a gold standard and was superimposed on the virtual head models, created by IPS CaseDesigner® and OrthoAnalyzer™, using an Iterative Closest Point algorithm. Differences between the positions of the intra-oral scans obtained with the software packages were recorded and expressed in six degrees of freedom as well as the inter- and intra-observer intra-class correlation coefficient.

RESULTS

The tested software packages, IPS CaseDesigner® and OrthoAnalyzer™, showed a high level of accuracy compared to the gold standard. The accuracy was calculated for all six degrees of freedom. It was noticeable that the accuracy in the cranial/caudal direction was the lowest for IPS CaseDesigner® and OrthoAnalyzer™ in both the maxilla and mandible. The inter- and intra-observer intra-class correlation coefficient showed a high level of agreement between the observers.

CLINICAL RELEVANCE

IPS CaseDesigner® and OrthoAnalyzer™ are reliable software packages providing an accurate fusion of the intra-oral scan in the CBCT. Both software packages can be used as an accurate fusion tool of the intra-oral scan in the CBCT which provides an accurate basis for 3D virtual planning.

The application and accuracy of feature matching on automated cephalometric superimposition

Background

The aim of this study was to establish a computer-aided automated method for cephalometric superimposition and to evaluate the accuracy of this method based on free-hand tracing.

Methods

Twenty-eight pairs of pre-treatment (T₁) and post-treatment (T₂) cephalograms were selected. Structural superimpositions of the anterior cranial base, maxilla and mandible were independently completed by three operators performing traditional hand tracing methods and by computerized automation using the feature matching algorithm. To quantitatively evaluate the differences between the two methods, the hand superimposed patterns were digitized. After automated and hand superimposition of T₂ cephalograms to T₁ cephalometric templates, landmark distances between paired automated and hand T₂ cephalometric landmarks were measured. Differences in hand superimposition among the operators were also calculated.

Results

The T₂ landmark differences in hand tracing between the operators ranged from 0.61 mm to 1.65 mm for the three types of superimposition. There were no significant differences in accuracy between hand and automated superimposition (p > 0.05).

Conclusions

Computer-aided cephalometric superimposition provides comparably accurate results to those of traditional hand tracing and will provide a powerful tool for academic research.