Accurate and early detection of apical lesions is an important factor for preventing pain and simultaneously improving treatment prognosis.17 When apical lesions cannot be diagnosed at the early stage, the lesions may grow progressively inducing pain, swelling, and bone loss around the apex, which ultimately leads to loss of tooth. Various imaging modalities can be employed for detecting apical lesions. Among them, the panoramic radiography is the most frequently used for identifying dental diseases at the first visit because the apical lesions can be detected at early stage using panoramic radiographs. However, radiographic detection of apical lesions is not objective between examiners and detecting ability is heavily dependent on the experience and skill of highly trained examiners.
To date, a few studies have employed deep learning as a tool for detecting apical lesions. A deep learning-based CNN algorithm enabled the automated detection of apical lesions efficiently and effectively with minimizing the dependence on the ability of examiners. However, to the best our knowledge, no study has examined the functionality of CNN for automated diagnosis of apical lesions thus far using entire panoramic radiographs. In this study, panoramic radiographs were used for training to detect apical lesions and the possibility of AI-guided diagnosis of apical lesions at early stage was confirmed.
U-Net and FPN models were used to detect apical lesions. Both algorithms are representative models for semantic segmentation and showed relatively great performance enough for clinical applications. U-Net demonstrated overall detection performance of 0.875 of precision, 0.855 of recall, 0.864 of F1-score while FPN showed 0.841 of precision, 0.865 of recall, 0.853 of F1-score, respectively.
Data augmentation is commonly used in training CNN models.15,18,19 It is an integral process of many state-of-the-art deep learning systems on image classification, object detection, and segmentation.20 Current deep neural networks have a number of parameters, tending to overfit the limited training data. Data augmentation is used to increase both the quantity and diversity of training data, thus preventing the overfitting and improving generalization.21 In this study, online data augmentation was used. It can optimize data augmentation and target network training in an online manner. The merits of online augmentation is the opposite features of offline methods.22 Their complementary character makes it possible to apply them together. The augmentation network can apply to the target network through online training from the start to finish excluding the inconveniences of pre-training or early stopping. Learning offline methods usually rely on distributed training, as there are many parallel optimization processes, but online data augmentation is simple and easy for training.
The performances were evaluated under even 0.01 of IoU value. In the most cases, the lesion area has a slight change in color or contrast in the panoramic radiographs and also can be observed as just a few pixels. Even if the lesion area and the CNN-predicted area match only a very small amount, it can be considered that the result is valid for this reason. Therefore, parameters for evaluating performances at the IoU value of 0.01 can be used as the performance index value of the CNN model in this study.
CNN-based models identified the apical lesions on maxilla and mandible with high performances, but it showed higher accuracy in the mandible than in the maxilla. Many anatomical structures such as sinus floor, nasal cavity, anterior nasal spine was superimposed and interfered the segmentation process of CNN. In contrast, there were not many overlapping anatomical structures on mandible, which indicated the better results.