Objectives: To retrospectively assess radiographic data and to prospectively classify radiographs (namely, panoramic, bitewing, periapical, and cephalometric images), we compared three deep learning architectures for their classification performance. Methods: Our dataset consisted of 31,288 panoramic, 43,598 periapical, 14,326 bitewing, and 1176 cephalometric radiographs from two centers (Berlin/Germany; Lucknow/India). For a subset of images L (32,381 images), image classifications were available and manually validated by an expert. The remaining subset of images U was iteratively annotated using active learning, with ResNet-34 being trained on L, least confidence informative sampling being performed on U, and the most uncertain image classifications from U being reviewed by a human expert and iteratively used for re-training. We then employed a baseline convolutional neural networks (CNN), a residual network (another ResNet-34, pretrained on ImageNet), and a capsule network (CapsNet) for classification. Early stopping was used to prevent overfitting. Evaluation of the model performances followed stratified k-fold cross-validation. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to provide visualizations of the weighted activations maps. Results: All three models showed high accuracy (>98%) with significantly higher accuracy, F1-score, precision, and sensitivity of ResNet than baseline CNN and CapsNet (p < 0.05). Specificity was not significantly different. ResNet achieved the best performance at small variance and fastest convergence. Misclassification was most common between bitewings and periapicals. For bitewings, model activation was most notable in the inter-arch space for periapicals interdentally, for panoramics on bony structures of maxilla and mandible, and for cephalometrics on the viscerocranium. Conclusions: Regardless of the models, high classification accuracies were achieved. Image features considered for classification were consistent with expert reasoning.

Objectives: The present study aimed to train deep convolutional neural networks (CNNs) to detect caries lesions on Near-Infrared Light Transillumination (NILT) imagery obtained either in vitro or in vivo and to assess the models’ generalizability. Methods: In vitro, 226 extracted posterior permanent human teeth were mounted in a diagnostic model in a dummy head. Then, NILT images were generated (DIAGNOcam, KaVo, Biberach), and images were segmented tooth-wise. In vivo, 1319 teeth from 56 patients were obtained and segmented similarly. Proximal caries lesions were annotated pixel-wise by three experienced dentists, reviewed by a fourth dentist, and then transformed into binary labels. We trained ResNet classification models on both in vivo and in vitro datasets and used 10-fold cross-validation for estimating the performance and generalizability of the models. We used GradCAM to increase explainability. Results: The tooth-level prevalence of caries lesions was 41% in vitro and 49% in vivo, respectively. Models trained and tested on in vivo data performed significantly better (mean ± SD accuracy: 0.78 ± 0.04) than those trained and tested on in vitro data (accuracy: 0.64 ± 0.15; p < 0.05). When tested in vitro, the models trained in vivo showed significantly lower accuracy (0.70 ± 0.01; p < 0.01). Similarly, when tested in vivo, models trained in vitro showed significantly lower accuracy (0.61 ± 0.04; p < 0.05). In both cases, this was due to decreases in sensitivity (by −27% for models trained in vivo and −10% for models trained in vitro). Conclusions: Using in vitro setups for generating NILT imagery and training CNNs comes with low accuracy and generalizability. Clinical significance: Studies employing in vitro imagery for developing deep learning models should be critically appraised for their generalizability. Applicable deep learning models for assessing NILT imagery should be trained on in vivo data.

The detection and classification of cystic lesions of the jaw is of high clinical relevance and represents a topic of interest in medical artificial intelligence research. The human clinical diagnostic reasoning process uses contextual information, including the spatial relation of the detected lesion to other anatomical structures, to establish a preliminary classification. Here, we aimed to emulate clinical diagnostic reasoning step by step by using a combined object detection and image segmentation approach on panoramic radiographs (OPGs). We used a multicenter training dataset of 855 OPGs (all positives) and an evaluation set of 384 OPGs (240 negatives). We further compared our models to an international human control group of ten dental professionals from seven countries. The object detection model achieved an average precision of 0.42 (intersection over union (IoU): 0.50, maximal detections: 100) and an average recall of 0.394 (IoU: 0.50–0.95, maximal detections: 100). The classification model achieved a sensitivity of 0.84 for odontogenic cysts and 0.56 for non-odontogenic cysts as well as a specificity of 0.59 for odontogenic cysts and 0.84 for non-odontogenic cysts (IoU: 0.30). The human control group achieved a sensitivity of 0.70 for odontogenic cysts, 0.44 for non-odontogenic cysts, and 0.56 for OPGs without cysts as well as a specificity of 0.62 for odontogenic cysts, 0.95 for non-odontogenic cysts, and 0.76 for OPGs without cysts. Taken together, our results show that a combined object detection and image segmentation approach is feasible in emulating the human clinical diagnostic reasoning process in classifying cystic lesions of the jaw.

We aimed to assess the effects of hyperparameter tuning and automatic image augmentation for deep learning-based classification of orthodontic photographs along the Angle classes. Our dataset consisted of 605 images of Angle class I, 1038 images of class II, and 408 images of class III. We trained ResNet architectures for classification of different combinations of learning rate and batch size. For the best combination, we compared the performance of models trained with and without automatic augmentation using 10-fold cross-validation. We used GradCAM to increase explainability, which can provide heat maps containing the salient areas relevant for the classification. The best combination of hyperparameters yielded a model with an accuracy of 0.63–0.64, F1-score 0.61–0.62, sensitivity 0.59–0.65, and specificity 0.80–0.81. For all metrics, it was apparent that there was an ideal corridor of batch size and learning rate combinations; smaller learning rates were associated with higher classification performance. Overall, the performance was highest for learning rates of around 1–3 × 10⁻⁶ and a batch size of eight, respectively. Additional automatic augmentation improved all metrics by 5–10% for all metrics. Misclassifications were most common between Angle classes I and II. GradCAM showed that the models employed features relevant for human classification, too. The choice of hyperparameters drastically affected the performance of deep learning models in orthodontics, and automatic image augmentation resulted in further improvements. Our models managed to classify the dental sagittal occlusion along Angle classes based on digital intraoral photos.