Deep chest X‐ray: Detection and classification of lesions based on deep convolutional neural networks. Issue 1 (21st October 2020)

• Record Type: Journal Article
• Authors: Cho, Yongwon; Lee, Sang Min; Cho, Young‐Hoon; Lee, June‐Goo; Park, Beomhee; Lee, Gaeun; Kim, Namkug; Seo, Joon Beom
• Is Part Of: International journal of imaging systems and technology. Volume 31:Issue 1(2021)
• Page Start: 72
• Page End: 81
• Publication Date: 2020-10-21
• Subjects: chest radiographs -- computer‐aided detection -- deep learning -- lung diseases -- machine learning -- radiography; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22508
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: We investigated whether a convolutional neural network (CNN) can enhance the usability of computer‐aided detection (CAD) of chest radiographs for various pulmonary abnormal lesions. The numbers of normal and abnormal patients were 6055 and 3463, respectively. Two radiologists delineated regions of interest for lesions and labeled the disease types as ground truths. The datasets were split into training, tuning, and testing as 7:1: 2. Total test sets were randomly selected in 1214 normal and 690 abnormal. A 5‐fold, cross‐validation was performed on our datasets. For the classification of normal and abnormal, we developed a CNN based on DenseNet169; for abnormal detection, The You Only Look Once (YOLO) v2 with DenseNet was used. Detection and classification of normal and five classes of diseases (nodule[s], consolidation, interstitial opacity, pleural effusion, and pneumothorax) on chest radiographs were analyzed. Our CNN model classified chest radiographs as normal or abnormal with an accuracy of 97.8%. For the results of the abnormal, F1 score, was 75.2 ± 2.28% for nodules, 55.0 ± 4.3% for consolidation, 78.2 ± 7.85% for interstitial opacity, 81.6 ± 2.07% for pleural effusion, and 70.0 ± 7.97% for pneumothorax, respectively. In addition, we conducted the experiments between our method and RetinaNet with only nodules. The results of our method and RetinaNet at cutoff‐0.5 in the free response operating characteristic curve were 83.45% and 80.55%, respectively. Our algorithm demonstrated viable detection and disease classification capacity and could be used for CAD of lung diseases on chest radiographs.