Improved Diagnostic Accuracy of Ameloblastoma and Odontogenic Keratocyst on Cone-Beam CT by Artificial Intelligence

Abstract

Objective: The purpose of this study was to utilize a convolutional neural network (CNN) to make preoperative differential diagnoses between ameloblastoma (AME) and odontogenic keratocyst (OKC) on cone-beam CT (CBCT).

Methods: The CBCT images of 178 AMEs and 172 OKCs were retrospectively retrieved from the Hospital of Stomatology, Wuhan University. The datasets were randomly split into a training dataset of 272 cases and a testing dataset of 78 cases. Slices comprising lesions were retained and then cropped to suitable patches for training. The Inception v3 deep learning algorithm was utilized, and its diagnostic performance was compared with that of oral and maxillofacial surgeons.

Results: The sensitivity, specificity, accuracy, and F1 score were 87.2%, 82.1%, 84.6%, and 85.0%, respectively. Furthermore, the average scores of the same indexes for 7 senior oral and maxillofacial surgeons were 60.0%, 71.4%, 65.7%, and 63.6%, respectively, and those of 30 junior oral and maxillofacial surgeons were 63.9%, 53.2%, 58.5%, and 60.7%, respectively.

Conclusion: The deep learning model was able to differentiate these two lesions with better diagnostic accuracy than clinical surgeons. The results indicate that the CNN may provide assistance for clinical diagnosis, especially for inexperienced surgeons.

Keywords: Inception v3; ameloblastoma; cone-beam CT; convolutional neural network; deep learning; odontogenic keratocyst.

Figure 1

(A) Ameloblastoma (AME). Axial view of CBCT shows the lesion with buccal expansion, obvious cortical bone resorption, and a multilocular pattern. (B) Typical H&E staining of AME (×200). (C) Odontogenic keratocyst (OKC). Axial view of CBCT shows that the lesion grows along the bone, with unapparent disruption of the cortical bone and the unilocular pattern. (D) Typical H&E staining of OKC (×200).

Figure 2

Flow diagram of the study. The training and testing datasets contained 272 and 78 patients, respectively. A total of 189 images comprising the region of interest (ROI) of one patient in the training dataset were cropped into smaller rectangles, padded into square images using black, and resized to 150 * 150 to gear the CNN. To reduce redundancy, we selected one image out of every three images. The series in the testing dataset underwent the same process except for changing the number of images. As shown in the picture, 120 slices of AME patient were tested by the trained CNN. Ninety slices were predicted to have AME and 30 slices were predicted to have OKC, so AME was ultimately considered.

Figure 3

The images used for training were obtained after a series of processing steps, and these 12 images came from one patient.

Figure 4

Inception v3 consists of five convolutional layers, two max-pooling layers, 11 inception modules, one average pooling layer, and one fully connected layer.

Figure 5

Convergence of the network training. At each epoch, the model was trained using all images in the training dataset, and the accuracy was evaluated. At the end of each epoch, we measured the accuracy of the model on the validation dataset. After 100 epochs, the training was stopped since both accuracy and cross-entropy loss would not be further improved.

Figure 6

Confusion matrices of Inception v3 and five oral and maxillofacial surgeons showed the specific diagnostic performance. The color shade of the grid represented the proportion of each class.