Young-Gon Kim1,2, Sang Min Lee3, Kyung Hee Lee4, Ryoungwoo Jang1,2, Joon Beom Seo5, Namkug Kim6. 1. Department of Biomedical Engineering, Asan Institute of Life Science, University of Ulsan College of Medicine, Asan Medical Center, Seoul, South Korea. 2. Department of Convergence Medicine, University of Ulsan College of Medicine, Asan Medical Center, Center 88, Olympic-ro 43-gil, Seoul, South Korea. 3. Department of Radiology, University of Ulsan College of Medicine, Asan Medical Center, 88 Olympic-ro 43 Gil, Songpa-gu, Seoul, South Korea. sangmin.lee.md@gmail.com. 4. Department of Radiology, Seoul National University Bundang Hospital, Seongnam, South Korea. 5. Department of Radiology, University of Ulsan College of Medicine, Asan Medical Center, 88 Olympic-ro 43 Gil, Songpa-gu, Seoul, South Korea. 6. Department of Convergence Medicine, University of Ulsan College of Medicine, Asan Medical Center, Center 88, Olympic-ro 43-gil, Seoul, South Korea. namkugkim@gmail.com.
Abstract
OBJECTIVES: To investigate the optimal input matrix size for deep learning-based computer-aided detection (CAD) of nodules and masses on chest radiographs. METHODS: We retrospectively collected 2088 abnormal (nodule/mass) and 352 normal chest radiographs from two institutions. Three thoracic radiologists drew 2758 abnormalities regions. A total of 1736 abnormal chest radiographs were used for training and tuning convolutional neural networks (CNNs). The remaining 352 abnormal and 352 normal chest radiographs were used as a test set. Two CNNs (Mask R-CNN and RetinaNet) were selected to validate the effects of the squared different matrix size of chest radiograph (256, 448, 896, 1344, and 1792). For comparison, figure of merit (FOM) of jackknife free-response receiver operating curve and sensitivity were obtained. RESULTS: In Mask R-CNN, matrix size 896 and 1344 achieved significantly higher FOM (0.869 and 0.856, respectively) for detecting abnormalities than 256, 448, and 1792 (0.667-0.820) (p < 0.05). In RetinaNet, matrix size 896 was significantly higher FOM (0.906) than others (0.329-0.832) (p < 0.05). For sensitivity of abnormalities, there was a tendency to increase sensitivity when lesion size increases. For small nodules (< 10 mm), the sensitivities were 0.418 and 0.409, whereas the sensitivities were 0.937 and 0.956 for masses. Matrix size 896 and 1344 in Mask R-CNN and matrix size 896 in RetinaNet showed significantly higher sensitivity than others (p < 0.05). CONCLUSIONS: Matrix size 896 had the highest performance for various sizes of abnormalities using different CNNs. The optimal matrix size of chest radiograph could improve CAD performance without additional training data. KEY POINTS: • Input matrix size significantly affected the performance of a deep learning-based CAD for detection of nodules or masses on chest radiographs. • The matrix size 896 showed the best performance in two different CNN detection models. • The optimal matrix size of chest radiographs could enhance CAD performance without additional training data.
OBJECTIVES: To investigate the optimal input matrix size for deep learning-based computer-aided detection (CAD) of nodules and masses on chest radiographs. METHODS: We retrospectively collected 2088 abnormal (nodule/mass) and 352 normal chest radiographs from two institutions. Three thoracic radiologists drew 2758 abnormalities regions. A total of 1736 abnormal chest radiographs were used for training and tuning convolutional neural networks (CNNs). The remaining 352 abnormal and 352 normal chest radiographs were used as a test set. Two CNNs (Mask R-CNN and RetinaNet) were selected to validate the effects of the squared different matrix size of chest radiograph (256, 448, 896, 1344, and 1792). For comparison, figure of merit (FOM) of jackknife free-response receiver operating curve and sensitivity were obtained. RESULTS: In Mask R-CNN, matrix size 896 and 1344 achieved significantly higher FOM (0.869 and 0.856, respectively) for detecting abnormalities than 256, 448, and 1792 (0.667-0.820) (p < 0.05). In RetinaNet, matrix size 896 was significantly higher FOM (0.906) than others (0.329-0.832) (p < 0.05). For sensitivity of abnormalities, there was a tendency to increase sensitivity when lesion size increases. For small nodules (< 10 mm), the sensitivities were 0.418 and 0.409, whereas the sensitivities were 0.937 and 0.956 for masses. Matrix size 896 and 1344 in Mask R-CNN and matrix size 896 in RetinaNet showed significantly higher sensitivity than others (p < 0.05). CONCLUSIONS: Matrix size 896 had the highest performance for various sizes of abnormalities using different CNNs. The optimal matrix size of chest radiograph could improve CAD performance without additional training data. KEY POINTS: • Input matrix size significantly affected the performance of a deep learning-based CAD for detection of nodules or masses on chest radiographs. • The matrix size 896 showed the best performance in two different CNN detection models. • The optimal matrix size of chest radiographs could enhance CAD performance without additional training data.
Keywords:
Algorithms; Computer-assisted diagnosis; Deep learning; Lung; Thoracic radiography
Authors: Yongwon Cho; Sung Ho Hwang; Yu-Whan Oh; Byung-Joo Ham; Min Ju Kim; Beom Jin Park Journal: Int J Imaging Syst Technol Date: 2021-05-13 Impact factor: 2.177