Non-Cochrane · DTA · PM41620618
Diagnostic accuracy of artificial intelligence models for temporomandibular joint anomalies on MRI: a systematic review and meta-analysis.
Pradhan Abhimanyu, Panda Aakash, Kadavigere Rajagopal, et al. · Biomedical engineering online · 2026 · PubMed · doi:10.1186/s12938-026-01525-6
Findings
Population Patients undergoing TMJ MRI for temporomandibular joint anomalies
22 findings
| AI (machine/deep learning) models, all architectures pooled | Reference standard diagnosis of TMJ anomaly | Specificity Human TMJ MRI datasets · All models | Specificity (pooled proportion) 0.399 [0.343, 0.456] GRADE not reported 5 studies |
| AI (machine/deep learning) models, all architectures pooled | Reference standard diagnosis of TMJ anomaly | Accuracy Human TMJ MRI datasets · All models Unresolved: the two extractors disagree on this row. | Accuracy (pooled proportion) 0.487 [0.403, 0.571] GRADE not reported 3 studies |
| AI (machine/deep learning) models, all architectures pooled | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · All models | Sensitivity (pooled proportion) 0.399 [0.348, 0.450] GRADE not reported 6 studies |
| ResNet-18 models | Reference standard diagnosis of TMJ anomaly | Specificity Human TMJ MRI datasets · Model architecture: ResNet 18 (Nozawa et al.) | Specificity (pooled proportion) 0.738 [0.693, 0.783] GRADE not reported 1 study |
| ResNet-18 models | Reference standard diagnosis of TMJ anomaly | Accuracy Human TMJ MRI datasets · Model architecture: ResNet-18 (Nozawa et al.) Unresolved: the two extractors disagree on this row. | Accuracy (pooled proportion) 0.732 [0.686, 0.778] GRADE not reported 1 study |
| ResNet-18 models | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: ResNet 18 (Nozawa et al.) | Sensitivity (pooled proportion) 0.724 [0.677, 0.770] GRADE not reported 1 study |
| Inception v3 models | Reference standard diagnosis of TMJ anomaly | Accuracy Human TMJ MRI datasets · Model architecture: Inception v3 (Nozawa et al.) Unresolved: the two extractors disagree on this row. | Accuracy (pooled proportion) 0.712 [0.666, 0.759] GRADE not reported 1 study |
| Inception v3 models | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: Inception v3 (Nozawa et al.) | Sensitivity (pooled proportion) 0.697 [0.624, 0.769] GRADE not reported 1 study |
| Inception v3 models | Reference standard diagnosis of TMJ anomaly | Specificity Human TMJ MRI datasets · Model architecture: Inception v3 (Nozawa et al.) | Specificity (pooled proportion) 0.725 [0.679, 0.771] GRADE not reported 1 study |
| EfficientNet b4 models | Reference standard diagnosis of TMJ anomaly | Specificity Human TMJ MRI datasets · Model architecture: EfficientNet b4 (Nozawa et al.) | Specificity (pooled proportion) 0.689 [0.642, 0.737] GRADE not reported 1 study |
| EfficientNet b4 models | Reference standard diagnosis of TMJ anomaly | Accuracy Human TMJ MRI datasets · Model architecture: EfficientNet b4 (Nozawa et al.) Unresolved: the two extractors disagree on this row. | Accuracy (pooled proportion) 0.701 [0.654, 0.748] GRADE not reported 1 study |
| EfficientNet b4 models | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: EfficientNet b4 (Nozawa et al.) | Sensitivity (pooled proportion) 0.714 [0.667, 0.760] GRADE not reported 1 study |
| GoogLeNet models | Reference standard diagnosis of TMJ anomaly | Accuracy Human TMJ MRI datasets · Model architecture: GoogLeNet (Nozawa et al.) Unresolved: the two extractors disagree on this row. | Accuracy (pooled proportion) 0.669 [0.621, 0.718] GRADE not reported 1 study |
| GoogLeNet models | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: GoogLeNet (Nozawa et al.) | Sensitivity (pooled proportion) 0.689 [0.641, 0.736] GRADE not reported 1 study |
| GoogLeNet models | Reference standard diagnosis of TMJ anomaly | Specificity Human TMJ MRI datasets · Model architecture: GoogLeNet (Nozawa et al.) | Specificity (pooled proportion) 0.649 [0.600, 0.698] GRADE not reported 1 study |
| ResNet-152 models | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: ResNet 152 (Li et al.) | Sensitivity (pooled proportion) 0.0819 [0.0722, 0.0915] GRADE not reported 1 study |
| ResNet-152 models | Reference standard diagnosis of TMJ anomaly | Accuracy Human TMJ MRI datasets · Model architecture: ResNet 152 (Li et al.) Unresolved: the two extractors disagree on this row. | Accuracy (pooled proportion) 0.0862 [0.0763, 0.0961] GRADE not reported 1 study |
| ResNet-152 models | Reference standard diagnosis of TMJ anomaly | Specificity Human TMJ MRI datasets · Model architecture: ResNet 152 (Li et al.) | Specificity (pooled proportion) 0.097 [0.086, 0.107] GRADE not reported 1 study |
| U-net family models (U-net, U-net++, U-net 3+) | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: U-net, U-net++, U-net 3+ (Min et al.) | Sensitivity (pooled proportion) 0.401 [0.353, 0.450] GRADE not reported 1 study |
| DeepLabV3+ and SegNet models | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: DeepLabV3+, SegNet (Min et al.) | Sensitivity (pooled proportion) 0.326 [0.283, 0.369] GRADE not reported 1 study |
| CNN fine-tuning and ensemble models | Reference standard diagnosis of TMJ anomaly | Sensitivity Human TMJ MRI datasets · Model architecture: CNN fine-tuning PDW, fine-tuning T2W, ensemble (Lee et al.) | Sensitivity (pooled proportion) 0.0375 [0.0318, 0.0432] GRADE not reported 1 study |
| CNN fine-tuning and ensemble models | Reference standard diagnosis of TMJ anomaly | Specificity Human TMJ MRI datasets · Model architecture: CNN fine-tuning PDW, T2W, ensemble (Lee et al.) | Specificity (pooled proportion) 0.058 [0.051, 0.065] GRADE not reported 1 study |
Included primary studies 14
Open in Studies| Study | Year | Also included by | Outcomes it informs here | Trial data |
|---|---|---|---|---|
| Nozawa 2024 Can temporomandibular joint osteoarthritis be diagnosed on MRI proton density-weighted images with diagnostic support from the latest deep… PMID:39067043 | 2025 | 1 other review | Accuracy; Sensitivity; Specificity | not extracted yet |
| Azma 2025 Automated pediatric TMJ articular disk identification and displacement classification in MRI with machine learning. PMID:39952550 | 2025 | this review only | not stated | not extracted yet |
| Ha 2025 Magnetic resonance image generation using enhanced TransUNet in temporomandibular disorder patients. PMID:40104864 | 2025 | this review only | not stated | not extracted yet |
| Kim 2021 Using deep learning to predict temporomandibular joint disc perforation based on magnetic resonance imaging. PMID:33758266 | 2021 | this review only | Sensitivity; Specificity | not extracted yet |
| Lee 2022 Advantages of deep learning with convolutional neural network in detecting disc displacement of the temporomandibular joint in magnetic… PMID:35790841 | 2022 | this review only | Accuracy; Sensitivity; Specificity | not extracted yet |
| Lee 2024 Automatic detection and visualization of temporomandibular joint effusion with deep neural network. PMID:39143180 | 2024 | this review only | Sensitivity; Specificity | not extracted yet |
| Lee 2025 (zero-TE MRI) Deep learning image enhancement for confident diagnosis of TMJ osteoarthritis in zero-TE MR imaging. PMID:39989448 | 2025 | this review only | not stated | not extracted yet |
| Li 2022 Temporomandibular joint segmentation in MRI images using deep learning. PMID:36368120 | 2022 | this review only | not stated | not extracted yet |
| Li 2023 Detecting anteriorly displaced temporomandibular joint discs using super-resolution magnetic resonance imaging: a multi-center study. PMID:38250655 | 2023 | this review only | Accuracy; Sensitivity; Specificity | not extracted yet |
| Min 2024 Enhanced multistage deep learning for diagnosing anterior disc displacement in the temporomandibular joint using MRI. PMID:39024472 | 2024 | this review only | Sensitivity | not extracted yet |
| Su 2025 Automatic classification of temporomandibular joint disorders by magnetic resonance imaging and convolutional neural networks. PMID:39873009 | 2025 | this review only | not stated | not extracted yet |
| Yoon 2023 Explainable deep learning-based clinical decision support engine for MRI-based automated diagnosis of temporomandibular joint anterior… PMID:36933315 | 2023 | this review only | not stated | not extracted yet |
| Yoon 2024 Multi-class segmentation of temporomandibular joint using ensemble deep learning. PMID:39160234 | 2024 | this review only | not stated | not extracted yet |
| Yoshimi 2024 Image preprocessing with contrast-limited adaptive histogram equalization improves the segmentation performance of deep learning for the… PMID:37263812 | 2024 | this review only | not stated | not extracted yet |
Overlap with other reviews
Reviews that include the same primary studies are not independent evidence. Shared counts match studies by PMID, DOI or NCT, and otherwise by author and year (a probable match). Only the non-Cochrane dental reviews have study lists, so overlap with Cochrane reviews is not counted yet.
| Review | Shared studies | Jaccard |
|---|---|---|
| Artificial Intelligence for Diagnosis of Temporomandibular and Cranio-Cervico-Mandibular Musculoskeletal Disorders: A Systematic Review… (2026) | 1 | 0.05 |
Review question
| Element | Planned (methods) | Found (results) |
|---|---|---|
| Population | Patients with TMJ disorders or mandibular canal abnormalities assessed with MRI (human MRI data) | 14 studies on TMJ MRI (South Korea, Japan, China, Canada, Taiwan); 6 studies in meta-analysis |
| Intervention | AI (machine learning or deep learning) models developed and tested on MRI for detection, classification, segmentation, treatment-response or prognosis of TMJ abnormalities | ResNet-18, Inception v3, EfficientNet b4, GoogLeNet, ResNet 152, U-net family, DeepLabV3+, SegNet, CNN fine-tuning, MLP, random forest, YOLO and others |
| Comparison | Not specified (reference standard not described) | Not specified |
| Outcomes | Accuracy, AUC, sensitivity, specificity, F1-score, Dice coefficient and related metrics on an independent test dataset | Pooled accuracy 0.487 (18 models), sensitivity 0.399 (29 models), specificity 0.399 (24 models) |
| Study designs | Original peer-reviewed retrospective or prospective observational and technical validation studies | 14 retrospective studies |
| Timing | Studies published January 2015 to September 2025 | Studies published 2021 to 2025 |