CBCT radiomics features combine machine learning to diagnose cystic lesions in the jaw.
Authors
Affiliations (5)
Affiliations (5)
- Department of Oral and Maxillofacial Radiology, School of Stomatology, Capital Medical University, Beijing 100070, China.
- Department of Clinical Research, SinoUnion Healthcare Inc, Beijing 100192, China.
- Department of Oral and Maxillofacial Pathology, School of Stomatology, Capital Medical University, Beijing 100070, China.
- Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing 100084, China.
- Department of Nuclear Medicine, Beijing Friendship Hospital, Capital Medical University, Beijing 100050, China.
Abstract
The aim of this study was to develop a radiomics model based on cone beam CT (CBCT) to differentiate odontogenic cysts (OCs), odontogenic keratocysts (OKCs), and ameloblastomas (ABs). In this retrospective study, CBCT images were collected from 300 patients diagnosed with OC, OKC, and AB who underwent histopathological diagnosis. These patients were randomly divided into training (70%) and test (30%) cohorts. Radiomics features were extracted from the images, and the optimal features were incorporated into random forest model, support vector classifier (SVC) model, logistic regression model, and a soft VotingClassifier based on the above 3 algorithms. The performance of the models was evaluated using a receiver operating characteristic (ROC) curve and the area under the curve (AUC). The optimal model among these was then used to establish the final radiomics prediction model, whose performance was evaluated using the sensitivity, accuracy, precision, specificity, and F1 score in both the training cohort and the test cohort. The 6 optimal radiomics features were incorporated into a soft VotingClassifier. Its performance was the best overall. The AUC values of the One-vs-Rest (OvR) multi-classification strategy were AB-vs-Rest 0.963; OKC-vs-Rest 0.928; OC-vs-Rest 0.919 in the training cohort and AB-vs-Rest 0.814; OKC-vs-Rest 0.781; OC-vs-Rest 0.849 in the test cohort. The overall accuracy of the model in the training cohort was 0.757, and in the test cohort was 0.711. The VotingClassifier model demonstrated the ability of the CBCT radiomics to distinguish the multiple types of diseases (OC, OKC, and AB) in the jaw and may have the potential to diagnose accurately under non-invasive conditions.