This study used a dataset of segmented panoramic radiographs. The reference test was the measured radiographically PBL (in % of the root length), quantified by three independent examiners. A deep CNN was constructed to detect PBL on the radiographs. We compared the performance of the CNNs against the subjective assessments by six dental practitioners. Reporting of this study follows the STARD guideline15.
Our primary performance metric was classification accuracy, which corresponds to the proportion of correct classifications per all classifications made. Secondary metrics were the AUC, which relates to a classifier’s (a model’s, an examiner’s) ability to avoid false classification, the F1-score which accounts for the relations between data’s positive labels and those given by a classifier, sensitivity and specificity, and the positive and negative predictive values. Details on the performance metrics are provided in the appendix.
Sample size calculation
Based on our hypothesis, we estimated the minimally required sample size allowing to detect significant differences in the accuracy between the two index tests, i.e. the CNN and the experienced dentists, when both assessed the same subjects (radiographs). Sample size calculation was based on the assumption that dentists would, in mean, show an accuracy of 0.80 to detect PBL11. We aimed to capture an accuracy-difference between dentists and the CNN of 0.03 (assuming the CNN to have an accuracy of 0.83 in mean) using a t-test (see below). Standard deviations were assumed to be 0.10 in both groups. At α = 0.05 and requiring an 80% power resulted in a sample size of n = 350 images (G* Power 220.127.116.11, Universität Düsseldorf, Germany). As we planned to use the majority of image segments for training, not validation, we eventually aimed to work with 1750 image segments; 350 for validation and 1400 for training.
We synthesized a dataset of 2001 manually cropped image segments, each focusing on one particular tooth, from 85 randomly chosen digital panoramic dental radiographs collected using Orthophos XG 3 (Sirona, Bensheim, Germany, year of construction: 2009) according to manufacturer’s instructions (considering patients sex and age etc.). Data collection was ethically approved (Charité ethics committee EA4/080/18); the ethics committee waivered the need for an informed consent given data being pseudonymized. Details on the imaging process can be found in the appendix. Only radiographs from dentate individuals were included. Prior to image processing, one examiner pre-screened the resulting tooth segments, and 264 tooth segments (mainly on incisors) were excluded, most of them as they heavily overlapped with the vertebrae and did not allow any kind of assessment (examples excluded and included images are provided in the appendix). No further in- or exclusion criteria were applied, i.e. the quality of the panoramic images (contrast, hazing, positioning etc.) was not used to exclude images (as we assumed a CNN needed to be able to detect PBL on possibly suboptimal images to be useful in the real world, or to highlight diagnostic uncertainty).
Our reference test was the maximal radiographically detectable PBL in % of the root length. For each tooth, three examiners independently and manually determined three points on each radiograph to estimate PBL in %; the mesial and distal cemento-enamel junction (CEJ), the deepest point of the root apex (for multi-rooted teeth, of the mesial and distal root) and the most apical extension of the alveolar crest (for multi-rooted teeth, the deepest extension mesial and distal was considered). If the CEJ was covered by a restoration, the most apical point of the restoration was used instead16. Using these points, it was now possible to calculate the % of PBL as the distance between the CEJ and the alveolar crest divided by the distance of the CEJ to the apex. For multi-rooted teeth, two % (one for the mesial and one for the distal root) were available and only the larger % recorded. Using the % PBL and not the absolute measures (in mm etc.) helps to overcome the described issue of patient positioning and magnification. The result of these three independent measurements were three %-value of PBL for each tooth segment (Fig. S2).
Details on these three measurements can be found in the appendix. In our base-case analysis, we first calculated the mean of these three measurements and then applied a cut-off value of 20%17 to distinguish between PBL being present (≥20%) or not (<20%). In a sensitivity analysis, we increased this cut-off to 25% and 30%, respectively, to account for the variability in PBL definitions, but also the different ease of detection of more or less severe PBL.
Both the measurement of the reference test as well as the examination of the radiographs by dentists (see below) were performed in Sidexis 4 (Sirona) using the length measurement tool. Measurements and examinations were performed in dimly lit rooms on diagnostic screens and standardized conditions. Both measurements and examinations allowed magnification and enhancement (contrast etc.) tools to be used.
Modelling via CNNs
A prominent use case of CNNs is to map an input (image) to an output (classification), based on a set of weights, learned from data. CNNs are composed of chained functions, often referred to as layers. Information is passed forward through the network layers to the final (output) layer, and thereby processed by applying intermediate computations18. CNNs are specialized kinds of neural networks, which use a mathematical operation called convolution that allows the CNN to extract features from image data. Stacked CNN layers identify different aspects of the input image such as edges, corners and spots or increasingly complex aspects of the image, such as shapes, structures and patterns. Finally, the last few network layers typically are able to take the feature-filtered images and translate them into votes, in our case a binary vote, of PBL being present or not.
The image data was digitally preprocessed: (1) Each image segment was transformed to gray-scale; (2) segments from the upper jaw were flipped by 180 degrees so that in all images, the crowns faced upwards and roots downwards; (3) all pixel values of each image segment were normalized to a fixed range [0, 1]; (4) all image segments we resized to 64 × 64 pixels. Further image augmentation techniques such as rotation, shearing and zooming were applied during CNN training. Data processing was performed using Python, and third party libraries such as NumPy, pandas, scikit-image and scikit-learn.
The dataset was randomly and repeatedly split into training and validation sets by applying group shuffling. An example can be found in the appendix (Table S2). 10-fold repetition of split and model training and validation was performed to evaluate the robustness of the CNN performance. Owing to the small data set size we did not evaluate the model on a hold-out test set. For each split, image segments of one panoramic radiograph (i.e., patient) were kept together either in the training set or in the validation set to prevent training and validation within the same patient. We further oversampled image instances from the minority class (in our case, these were images where PBL was present according to the reference test) to reduce the detrimental effect of class imbalance on model performance19. Accordingly, the prevalence of positive class in the training set was close to 0.5. We did not oversample the minority class during validation (Table 1).
CNNs were developed using the TensorFlow framework and Keras. We combined convolutional layers with activation functions (rectified linear units) and max-pooling layers. For the purpose of model regularization we interlaced these sequences with batch-normalization20 and dropout layers. As final model layers we used a series of fully connected dense layers and a softmax classifier. Details on the model development are provided in the appendix.
Hyperparameters were systematically tuned via grid search21. We considered the number and ordering of stacked layers, the number of units in the hidden layers, the number of convolutional filters, the kernel sizes and activation functions, image preprocessing, different types of optimizers and learning rates, the usage of dropout and batch normalization layers and their parameterizations and positions, the batch size, and image augmentation parameters as hyperparameters. Details on the hyperparameter tuning and image augmentation can be found in the appendix.
The performance of the CNNs against a reference test is only limitedly useful for interpretation. Hence, six experienced dentists (mean (SD; range) clinical experience 6 (3; 3–10 years) additionally assessed the dataset for radiographically detectable PBL (binary outcome). The dentists were either specialists in operative dentistry and periodontology (n = 1), specialized in endodontics (n = 1) or general dentists (n = 4). All worked full-time at a university hospital; three of them had worked in private practice before. Dentists were informed about the background of the study and the diagnostic task, and instructed to decide if, according to their professional opinion, PBL was present or not. We computed the operating point (sensitivity vs. 1-specificity) for each dentist in order to evaluate the examiner’s discrimination ability, and calculated a mean (SD) AUC22. A one-sided two-sample Welch’s t-test was used to compare the dentists’ and the CNN’s accuracy, with α = 0.05 as level of significance. In order to assess the inter-rater reliability we computed Fleiss kappa23, assuming 0–0.20 as slight, 0.21–0.40 as fair, 0.41–0.60 as moderate, 0.61–0.80 as substantial, and 0.81–1 as almost perfect agreement24. Dentists did not revisit their records, and intra-rater reliability was not assessed.
A range of sensitivity analyses were performed. First, we applied the reference test cut-offs at 25% and 30% PBL. Second, we systematically evaluated the model’s discrimination ability on different tooth types (molars, premolars, canines, incisors), the rationale being that owing to the radiographic image generation process, anterior teeth are usually more difficult to assess than posterior teeth.
Ethical approval and informed consent
All experiments were carried out in accordance with relevant guidelines and regulations. Data collection was ethically approved (Charité ethics committee EA4/080/18).