Computer aided diagnosis system for cervical lymph nodes in CT images using deep learning. (January 2022)

• Record Type: Journal Article
• Authors: Tekchandani, Hitesh; Verma, Shrish; Londhe, Narendra D.; Jain, Rajiv Ratan; Tiwari, Avani
• Is Part Of: Biomedical signal processing and control. Volume 71(2022)Part A
• Publication Date: 2022-01
• Subjects: Lymph node -- Malignant -- Benign -- Computed tomography -- Deep learning -- Attention -- Squeeze and excitation; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103158
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Deep Learning based CADx system for cervical lymph nodes malignancy detection. Attention mechanism and residual concept based proposed LNdtnNet for detection of cervical lymph nodes in CT images. Modified VGG, modified Squeeze and Excitation, and residual concept based proposed LNdgsNet for the diagnosis of cervical lymph nodes in CT images. For this proposed study data is collected from Regional Cancer Center of Raipur Chhattisgarh, India. Abstract: Background and objective: Efficient treatment of head and neck cancer requires fast and reliable detection and diagnosis of cervical lymph nodes (CLNs). In current practices, manual methods for detection and invasive onco-pathological tests for diagnosis are considered as the gold standards. These methods suffers from numerous shortcomings which makes them inefficient. This raises the need of a non-invasive and automated computer aided diagnosis (CADx) system. Such CADx system undermines the data for extracting the discriminant information and computed tomography (CT) images are information rich and non-invasive imaging modality for oncological diseases. The design of reliable CADx system demands both accurate detection and classification of CLNs in CT images. Methods: The authors have proposed the deep learning based innovative and customized architecture based on attention mechanism and residual concept with the base UNet model, for the CLNs detection part (LNdtnNet) of the CADx system. While another architecture based on squeeze and excitation network and residual network with the base model of modified VGG, is proposed for the remaining diagnosis part (LNdgsNet) of the proposed CADx System. Results: In first stage, the proposed LNdtnNet for CLNs detection found the best results of sensitivity = 92.78%, and Dice score = 94.18%. In second stage, proposed LNdgsNet attaining an average sensitivity, specificity, accuracy, and area under the curve of 95.62%, 93.88%, 95.28%, and 94.75%, respectively. Conclusion: The proposed both architectures trained offline run on a single platform back to back for testing cases. The overall results confirm the utility of the proposed CADx system.