Convolutional neural network analysis of cervical CT images: classification of Kikuchi disease, lymphoma, lymphadenitis, and tuberculosis.
Authors
Affiliations (6)
Affiliations (6)
- Department of Biomedical Engineering, Pusan National University Yangsan Hospitial, Yangsan, Korea.
- Department of Otolaryngology-Head and Neck Surgery, School of Medicine, Pusan National University, Yangsan, Korea. [email protected].
- Research Institute for Convergence of Biomedical Science and Technology, Pusan National University Yangsan Hospital, Yangsan, Korea. [email protected].
- Department of Biomedical Engineering, Pusan National University Yangsan Hospitial, Yangsan, Korea. [email protected].
- Research Institute for Convergence of Biomedical Science and Technology, Pusan National University Yangsan Hospital, Yangsan, Korea. [email protected].
- Department of Biomedical Engineering, School of Medicine, Pusan National University, Yangsan, Korea. [email protected].
Abstract
Kikuchi disease, lymphoma, lymphadenitis, and tuberculosis are common diseases affecting the head and neck. The causes and treatment methods differ; however, the initial symptoms of these diseases (fever, pain, and neck swelling) are generally similar; therefore, it is important to accurately determine the type of disease at its initial stage. During the performance evaluation, the values of precision-recall area under the curve (PR AUC) were 0.785, 0.731, 0.920, and 0.821 for the four single-disease detection models for Kikuchi disease, lymphoma, lymphadenitis, and tuberculosis, respectively; 0.819 for a model for classifying the type of treatment; and 0.807 for a model for classifying all diseases with a single inspection. In the model performance tests, all six implemented models showed relatively high performance for the test dataset (accuracy: 0.6864-0.9278, precision: 0.7532-0.9583, recall: 0.3896-0.8804, and F1-score: 0.5310-0.8710). Based on these experimental results, we conclude that the proposed CNN-based diagnosis support technique has the potential to be an efficient pre-screening tool for first-time patients with Kikuchi disease, lymphoma, lymphadenitis, and tuberculosis by reducing the dependency on high-risk invasive diagnosis; however, additional model enhancement is required to improve its clinical applicability.