Periodontitis bone loss detection in panoramic radiographs using modified YOLOv7

Datasets

The lack of publicly available periodontitis datasets has been one of the main constraints to conduct research in this pertinent field. To solve this problem, researchers from Kong et al. (2023) collaborated with dentists from Wuhan University’s Hospital of Stomatology to develop a panoramic radiography periodontitis dataset. Panoramic radiographs from the dental hospital’s electronic medical records were screened using the OC200D technology from Instrumentarium Dental, Inc. This dataset includes a wide range of clinically specific circumstances, i.e., missing teeth and teeth that have undergone root canal therapy, and consists of 1,747 panoramic images with a resolution of 1,536 × 2,976. It is significant to emphasize that the dataset does not contain any personal information and is just meant for academic research. Some sample Images from the dataset are shown in Fig. 1.

Dental professionals used the VGG Image Annotator (VIA) to annotate the periodontitis dataset used in this investigation (Dutta & Zisserman, 2019). It is noteworthy that only experts medical professionals can delineate the panoramic image with precision for periodontitis detection. Each tooth’s index, polygon mask, root bone loss (RBL) (four classes), and FI (three classes) are all included in the annotated information. It is important to note that even among the expert dentists, the detection of periodontitis in panoramic radiographs might be partially subjective and may not be entirely correct because of the inter-observer variability. It is also important to highlight that wisdom teeth that have not yet erupted are not marked in the dataset because they would not have periodontal disease. An example of a typical RBL annotation is shown in Fig. 2.

The panoramic radiographs dataset used in this study comprises 1,747 sample images, which contain a total of 50,080 tooth annotations. Detailed information about the radiographic bone loss (RBL) and furcation involvement (FI) annotations is provided in Tables 2 and 3, respectively. The annotation methodology adheres to the latest standard (Tonetti, Greenwell & Kornman, 2018), which defines four classes of tooth RBL and three classes of tooth FI. The ‘Healthy’ category is the most common in RBL, with a ratio of 0.403, while the ‘Severe’ category has the lowest ratio of 0.059. Similarly, in FI, the ‘Healthy’ category is the most common with a ratio of 0.929, and the ‘Severe’ category has the lowest ratio of 0.020. It is noteworthy that the number of annotations in each category is imbalanced.

YOLOv7-M

A state-of-the-art real-time object identification technology called YOLOv7 was recently released. One CNN is used, and it has been trained to recognise objects in both still and moving pictures. Compared to past iterations of YOLO (v1, v2, v3, v4, v5, v6), YOLOv7 introduced several modifications. To increase the accuracy of object detection, it combines anchor boxes, anchor box scaling, and anchor box ratios (Yang, Zhang & Liu, 2022). It also introduces new features such as mosaic data augmentation and self-adversarial training to improve generalization and robustness (Wang, Bochkovskiy & Liao, 2022). YOLOv7 is fast and accurate, making it a popular choice for a wide range of applications, including autonomous vehicles, surveillance, and robotics.

In this work, we propose a modified YOLOv7 model for the panoramic radiographs periodontitis detection model. The SpacetoDepth concept (Ridnik et al., 2021) is employed in the YOLO Focus module. Stacking the image’s quarters is intended to convert spatial data into depth data. The Focus module divides the image into different slices using a slicing operation. Concatenation is used to connect the slices of varying depths, which are then passed to the convolution layer. At the start of the network, we added a Focus module. The primary purpose of adding the Focus layer (Yan et al., 2021) at the start was to reduce the number of layers, parameters, FLOPS, while increasing training speed and CUDA memory and boosting forward and backward speed with minimal impact on mAP.

Additionally, the C3Ghost module is also used to replace the backbone C3 module. The C3 module comprises 3*Convolution and the CSP bottleneck (Wang et al., 2020), while the C3Ghost module comprises the C3 module and the Ghost bottleneck (Han et al., 2020). The Ghost module uses low-cost processes to generate extra feature maps. It is a plug-and-play component used to improve convolutional neural networks. Ghost bottlenecks are designed to stack Ghost modules, and GhostNet is easy to set up. In terms of recognition tasks, GhostNet can outperform. Finally, for faster convergence and more accurate results, we changed the activation function to ELU (Heusel et al., 2015). Figure 3 shows the modified Yolo7 model.

We observed a significant improvement in the performance through our proposed method. We began with the focus module, which enabled us to segment and down-sample the images without losing any information. Later, for the recognition task, the C3GhostNet module, which is a C3 module with a Ghost Bottleneck, was used instead of the C3 backbone.

Evaluation metrics

In this work, we evaluated the performance of the YOLOV7-M using four performance metrics: precision (P), recall (R), mean average precision (mAP), and F1-score. Precision (P), which measures how successfully a model can categorize an object, is the percentage of correctly recognized samples out of all samples that were detected. Recall (R) measures how well the model was able to identify the object by expressing the proportion of samples that were successfully detected out of all real samples. mAP is the average AP across all categories, reflecting the overall performance of the model, while AP represents the average precision at various recall rates. Model performance is positively correlated with the F1-score, which is calculated by combining the precision and recall values. The formulae are as follows:

(1) $$P=\frac{TP}{TP+FN}$$

(2) $$R=\frac{TP}{TP+FN}$$

(3) $$AP={\int }_0^{1}P(R)dR$$

(4) $$mAP=\frac{1}{Q}\sum _{q\in Q}AP(q)$$

(5) $$F1=\frac{2\times P\times R}{P+R}$$where TP is the number of positive samples the model anticipated as being positive, FP is the number of negative samples the model forecasted as being positive, and FN is the number of positive samples the model forecasted as being negative (Naseer et al., 2022).

In addition to model architecture and before start training the YOLOv7, we need a loss function to update the model parameters. Since object detection is a complex problem to teach a model, the loss functions of such models are usually quite complex, and YOLOv7 is not an exception (Wang, Bochkovskiy & Liao, 2022). To make the problem computationally cheaper, the YOLOv7 first finds the anchor boxes that are likely to match each target box and treat them differently, known as the center prior anchor boxes. This process is applied at each FPN head, for each target box, across all images in batch at once. The main idea behind FPNs is to leverage the nature of convolutional layers, which reduce the size of the feature space and increase the coverage of each feature in the initial image, to output predictions at different scales (Hughes & Camps, 2022). FPNs are usually implemented as a stack of convolutional layers, as we can see by inspecting the detection head of our YOLOv7 model. Figure 4 illustrates loss functions for Modified YOLOv7 for tooth detection to facilitate understanding. Figure 5 shows precision, recall and mAP, and Fig. 6 shows confusion matrices. The model was trained with 100 epochs and a batch size of 32.

The F1-score curve, also known as F-measure, is a widely used error metric to evaluate the performance of classification models. It ranges from 0 to 1, with 0 indicating the worst possible and 1 indicating the best possible. The F1-score considers both precision and recall, making it a robust metric for imbalanced datasets (Lipton, Elkan & Narayanaswamy, 2014). Figure 7 shows the F1-score curve for the YOLOv7-M model, which achieved an F1-score of 0.73 for tooth detection at a threshold of 0.405. This score indicates that the model performed well in classifying the four tooth classes.

When assessing prediction models, accuracy and recall are essential metrics to consider. Precision evaluates the relevance of predicted positive outcomes, while recall measures the model’s ability to predict positive samples, i.e., it assesses the model’s ability to identify all positive samples. A high precision indicates that the model predicts most positive samples accurately, whereas a high recall indicates that the model identifies most of the positive samples in the dataset. Precision-recall (PR) curves plot these metrics on the same graph, as illustrated in Fig. 8, and are commonly used to evaluate classification models. PR curves enable users to predict future classification results accurately by measuring the proportion of positive predictions that are true positives (Saito & Rehmsmeier, 2015). In the PR space, a predictor’s score closer to the perfect classification point (1,1) indicates better performance, whereas a score closer to zero implies worse performance. Figure 8 depicts the model’s performance on tooth detection in terms of the precision-recall measure.

The p-curve is a tool that can be used to evaluate the evidential value of research findings and rule out potential biases such as p-hacking and file-drawing (Simonsohn, Nelson & Simmons, 2019). It involves analyzing the distribution of statistically significant p-values (p $\le$ 0.05) across a set of studies that test the same hypotheses of interest. The assumption is that all significant parameters are included (Simonsohn, Simmons & Nelson, 2015). p-values indicate the likelihood of obtaining a particular result under the assumption that the null hypothesis is true. In Fig. 9, the p-curves for YOLOv7-M for tooth detection are displayed, showing the confidence level for precision and recall of 1 at a threshold of 0.952, indicating the accuracy of tooth detection.

The R-curve, also known as the recall curve, is a graphical representation of the trade-off between the recall and confidence levels in a classification model. It is constructed by plotting the recall (true positive rate) against different confidence thresholds. The recall indicates the proportion of positive samples that the model correctly identifies, while the confidence threshold is the minimum confidence level required for a prediction to be considered positive.

In the context of tooth periodontal bone loss detection, the R-curve shows how well the YOLOv7-M model performs in identifying teeth with different levels of RBL at different confidence levels. A high recall score at a given confidence threshold means that the model can detect a high percentage of positive samples (teeth with RBL) while keeping the false positive rate low. However, a low recall score means that the model misses a significant number of positive samples. By examining the R-curve, we can determine the optimal confidence threshold that achieves a balance between recall and precision for a specific task. In the case of tooth periodontal bone loss detection, we can use the R-curve to identify the threshold that maximizes the recall while maintaining an acceptable level of false positives, as shown in Fig. 10.

Table 5 shows the experimental results of the proposed YOLOv7-M model in detecting periodontitis bone loss in panoramic radiographs. The dataset is split into three subsets: 70% for training, 10% for validation, and 20% for testing. This 7: 1: 2 ratio ensures effective model learning, tuning, and evaluation. The model was evaluated using various evaluation measures, such as precision, recall, and [email protected], for different classes of severity of bone loss—healthy, medium, mild and severe.

As a result, the suggested model obtained for all classes combined a precision of 0.7, recall of 0.804, and [email protected] of 0.81. Five groups of severity of bone loss were found to have variable precision and recall rates. The suggested model had the highest precision (0.869) and recall (0.933) for the healthy class, showing that it could reliably identify healthy teeth. The model’s precision for the medium class was 0.598, and its recall was 0.728, indicating that it did a good job of spotting medium bone loss. The suggested model achieved precision values for the mild and severe classes of 0.653 and 0.68, respectively, and recall values of 0.863 and 0.693, respectively. These findings suggest that the suggested approach could, to a lesser extent than healthy teeth, detect mild and severe bone loss. The model achieved high precision and recall rates for healthy teeth and moderate performance in detecting medium bone loss. The performance of the model in detecting mild and severe bone loss was relatively lower but still showed potential for improvement. The results of this study can help to develop automated systems for the early detection and diagnosis of periodontitis, which can improve patient outcomes and reduce healthcare costs.

The five-fold cross-validation results for YOLO-based teeth segmentation across four classes (Healthy, Medium, Mild, Severe) presented in Table 6 demonstrate robust and consistent performance. The Mild class consistently outperformed others, achieving the highest average metrics: precision (0.933), recall (0.892), F1-score (0.937), and mAP (0.920). The Healthy class followed with averages of precision (0.917), recall (0.869), F1-score (0.928), and mAP (0.910). The Severe class had averages of precision (0.916), recall (0.863), F1-score (0.923), and mAP (0.905), while the medium class averaged precision (0.904), recall (0.875), F1-score (0.915), and mAP (0.904). Overall, the model achieved an average precision of 0.917, recall of 0.871, F1-score of 0.925, and mAP of 0.910 across all folds, highlighting its reliability and effectiveness in dental image segmentation, particularly for subtle conditions.

Comparison analysis of proposed model performance with other YOLO models

The performance of several YOLO models using panoramic radiographs derived from the periodontitis dataset are compared in this section. The model was trained for 100 epochs with a batch size of 32. We set the learning rate at 0.001 and utilized the SGD optimizer, which is well-suited for handling sparse gradients on noisy problems. The loss function employed was cross-entropy loss, which is appropriate for our classification tasks and have been implemented in several other applications (Waqas et al., 2024; Hanif et al., 2023). To ensure compatibility with our experimental environment, the images were resized to 640 $\times$ 640 pixels. This size was chosen to balance computational efficiency and model performance. Additional hyperparameters included a weight decay of 0.0005 to prevent overfitting and a momentum of 0.9 to accelerate the convergence of the model. A comparison of the proposed YOLOv7-M performance model’s accuracy, recall, F1-score, mAP, and training time with other significant YOLO versions is shown in Table 7. Precision is 0.695, recall 0.784, F1-score 0.726, and mAP is 0.748 for the YOLO-v4 model. YOLO-v4 training takes 1 h and 57 min to complete. Accuracy of 0.706, recall of 0.786, F1-score of 0.706, and mAP of 0.793 are displayed by YOLOv5. YOLOv5 training takes 1 h and 43 min. Precision 0.721, recall 0.8, F1-score 0.874, and mAP is 0.829 for YOLOv7. YOLOv7 training takes an hour and a quarter. With an accuracy of 0.917, recall of 0.871, F1-score of 0.935, and mAP of 0.93, the suggested YOLOv7-M model performs better than other models in comparison. For YOLOv7-M, training takes an hour and fifteen minutes. YOLOv7-M performs better than the previous YOLO versions in terms of accuracy, recall, F1-score, and mAP, as shown in Table 7. This shows that YOLOv7-M performs more accurately and robustly when identifying signs of periodontitis in panoramic radiographs.

In our experiments, the modified YOLOv7-M model demonstrated a significant improvement in frame rate performance, achieving an impressive 100–120 frames per second (FPS). This high FPS is critical for real-time applications, ensuring that the model can process panoramic radiographs swiftly and efficiently. The increased speed is attributed to the incorporation of the focus module and feature fusion module, which enhance the model’s inference capabilities while maintaining high accuracy. This superior performance positions YOLOv7-M as a highly effective tool for automated periodontitis diagnosis, capable of providing rapid and reliable assessments to assist dental professionals in timely and accurate detection of bone loss due to periodontitis.

Additionally, the proposed model achieves these superior results while requiring less training time compared to YOLO-v4, YOLOv5, and YOLOv7. The Focus module and the C3Ghost module in YOLOv7-M helped to improve the speed and precision. Ghost module also could compute fewer intrinsic maps and perform speedy linear operations on them, allowing it to work quicker than others. The same improvement for YOLOv7-M can be seen and this proved that currently, state-of-the-art real-time object detectors are mainly based on YOLO (Wang, Bochkovskiy & Liao, 2022; Redmon et al., 2016; Redmon & Farhadi, 2017, 2018). Finally, the proposed YOLOv7 surpasses all known object detectors in both speed and accuracy. These findings highlight the efficacy of the YOLOv7-M model as a promising approach for automated periodontitis detection on panoramic radiographs, exhibiting high accuracy and efficiency.

The evaluation based on the mean average precision (mAP) scores provided in Table 8. The mAP scores serve as a measure of the detection performance of the respective YOLO models in identifying periodontitis-related features in panoramic radiographs. Upon analyzing the results, it can be observed that YOLO-v7-M achieves the highest mAP score of 0.910, indicating its superior performance in detecting periodontitis-related features. This model demonstrates a remarkable capability to accurately identify and localize periodontal abnormalities within panoramic radiographs. Following YOLO-v7-M, YOLOv7 exhibits a high mAP score of 0.829, showcasing its effectiveness in periodontitis detection.

The mAP score of 0.793 for YOLOv5 is significantly lower, showing an opportunity for development. Compared to the other YOLO versions examined in this work, this model obtains a somewhat lower detection accuracy. In conclusion, the comparison study demonstrates that YOLO-v7-M achieves highest mAP score among the investigated models for periodontitis identification on panoramic radiographs. Despite having significantly lower mAP scores, YOLOv7, panoramic dental convolutional neural network (PDCNN), YOLO-v4, and Faster R-CNN also exhibit competitive performance. YOLOv5, although showing less accuracy, might still use further improvement to improve its detecting skills.

We provide the visual detection outcomes of all techniques in Fig. 11 to provide a more thorough knowledge of each network’s performance. These two images show a portion of the panoramic radiographs selected for the test dataset. All networks show the capacity to forecast the general location and class of the majority of teeth, in particular the root bone loss (RBL) class. However, YOLO-v4 and YOLO-v5 erroneously identify minor sections of tooth positions that are unrelated to periodontitis, leading to the production of duplicate forecasts on occasion. However, with few faults, the findings produced by PDCNN, YOLO-v7, and YOLOv7-M closely reflect the ground truth (GT). It is difficult to see substantial categorization differences visually, because there are not many examples available and these approaches all perform with around the same accuracy. These results offer insightful information on the effectiveness of YOLO models in automated periodontitis assessment on panoramic radiographs, increasing computer-aided detection systems in dentistry. These results can be used to improve the precision and effectiveness of periodontal detection and treatment planning through more study and optimization efforts.