TBNet: a context-aware graph network for tuberculosis diagnosis. (February 2022)

• Record Type: Journal Article
• Authors: Lu, Si-Yuan; Wang, Shui-Hua; Zhang, Xin; Zhang, Yu-Dong
• Is Part Of: Computer methods and programs in biomedicine. Volume 214(2022)
• Publication Date: 2022-02
• Subjects: tuberculosis -- computed tomography -- computer-aided diagnosis -- graph neural network -- random vector functional-link net; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.106587
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: We proposed a new deep model called TBNet for the diagnosis of TB in CT images. The sample-level features were extracted by transfer learning of EfficientNet. Context-aware features were proposed to use context information in feature space. A CARVFLN was proposed as the classifier in the proposed TBNet. Experimental results revealed that our TBNet achieved high classification performance. Abstract: Background and objective: Tuberculosis (TB) is an infectious bacterial disease. It can affect the human lungs, brain, bones, and kidneys. Pulmonary tuberculosis is the most common. This airborne bacterium can be transmitted with the droplets by coughing and sneezing. So far, the most convenient and effective method for diagnosing TB is through medical imaging. Computed tomography (CT) is the first choice for lung imaging in clinics because the conditions of the lungs can be interpreted from CT images. However, manual screening poses an enormous burden for radiologists, resulting in high inter-observer variances. Hence, developing computer-aided diagnosis systems to implement automatic TB diagnosis is an emergent and significant task for researchers and practitioners. This paper proposed a novel context-aware graph neural network called TBNet to detect TB from chest CT images Methods: Traditional convolutional neural networks can extract high-level image features to achieve good classification performance on the ImageNet dataset. However, we observed that the spatial relationships between the feature vectors are beneficial for the classification because the feature vector may share some common characteristics with its neighboring feature vectors. To utilize this context information for the classification of chest CT images, we proposed to use a feature graph to generate context-aware features. Finally, a context-aware random vector functional-link net served as the classifier of the TBNet to identify these context-aware features as TB or normal Results: The proposed TBNet produced state-of-the-art classification performance for detecting TB from healthy samples in the experiments Conclusions: Our TBNet can be an accurate and effective verification tool for manual screening in clinical diagnosis.