Y Ohta1, H Matsuzawa2, K Yamamoto3, Y Enchi2, T Kobayashi3, T Ishida4. 1. MedCity21, Division of Premier Preventive Medicine, Osaka City University Hospital, Abeno Harukasu 21F, Abenosuji 1-1-43, Abeno-ku Osaka, Osaka 545-8545, Japan. 2. Department of Radiology, Osaka University Hospital, Yamadaoka 2-15, Suita, Osaka 565-0871, Japan. 3. Department of Medical Physics and Engineering, Graduate School of Medicine, Osaka University, Yamadaoka 1-7, Suita, Osaka 565-0871, Japan. 4. Department of Medical Physics and Engineering, Graduate School of Medicine, Osaka University, Yamadaoka 1-7, Suita, Osaka 565-0871, Japan. Electronic address: tishida@sahs.med.osaka-u.ac.jp.
Abstract
INTRODUCTION: Lateral radiography of the knee joint is frequently performed; however, the retake rate is high owing to positioning errors. Therefore, in this study, to reduce the required number and time of image retakes, we developed a system that can classify the tilting directions of lateral knee radiographs and evaluated the accuracy of the proposed method. METHODS: Using our system, the tilting directions of a lateral knee radiographs were classified into four direction categories. The system was developed by training the DCNN based on 50 cases of Raysum images and tested on three types test dataset; ten more cases of Raysum images, one case of flexed knee joint phantom images and 14 rejected knee joint radiographs. To train a deep convolutional neural network (DCNN), we employed Raysum images created via three-dimensional (3D) X-ray computed tomography (CT); 11 520 Raysum images were created from 60 cases of 3D CT data by changing the projection angles. Thereby, we obtained pseudo images attached with correct labels that are essential for training. RESULTS: The overall accuracy on each test dataset was 88.5 ± 7.0% (mean ± standard deviation), 81.4 ± 11.2%, and 73.3 ± 9.2%. The larger the tilting degree of the knee joint, the higher the classification accuracy. CONCLUSION: DCNN could classify the tilting directions of a knee joint from lateral knee radiographs. Using Raysum images made it possible to facilitate creating dataset for training DCNN. The possibility was indicated for using support system of lateral knee radiographs. IMPLICATIONS FOR PRACTICE: The system may also reduce the burden on patients and increase the work efficiency of radiological technologists.
INTRODUCTION: Lateral radiography of the knee joint is frequently performed; however, the retake rate is high owing to positioning errors. Therefore, in this study, to reduce the required number and time of image retakes, we developed a system that can classify the tilting directions of lateral knee radiographs and evaluated the accuracy of the proposed method. METHODS: Using our system, the tilting directions of a lateral knee radiographs were classified into four direction categories. The system was developed by training the DCNN based on 50 cases of Raysum images and tested on three types test dataset; ten more cases of Raysum images, one case of flexed knee joint phantom images and 14 rejected knee joint radiographs. To train a deep convolutional neural network (DCNN), we employed Raysum images created via three-dimensional (3D) X-ray computed tomography (CT); 11 520 Raysum images were created from 60 cases of 3D CT data by changing the projection angles. Thereby, we obtained pseudo images attached with correct labels that are essential for training. RESULTS: The overall accuracy on each test dataset was 88.5 ± 7.0% (mean ± standard deviation), 81.4 ± 11.2%, and 73.3 ± 9.2%. The larger the tilting degree of the knee joint, the higher the classification accuracy. CONCLUSION:DCNN could classify the tilting directions of a knee joint from lateral knee radiographs. Using Raysum images made it possible to facilitate creating dataset for training DCNN. The possibility was indicated for using support system of lateral knee radiographs. IMPLICATIONS FOR PRACTICE: The system may also reduce the burden on patients and increase the work efficiency of radiological technologists.