Automatic CT Quantification of Coronavirus Disease 2019 pneumonia:
An international collaborative development, validation, and clinical
implication
Seung-Jin Yoo, Xiaolong Qi, Shohei Inui, Sang Joon Park, Hyungjin
Kim, Yeon Joo Jeong, Kyung Hee Lee, Young Kyung Lee, Bae Young Lee, Jin
Yong Kim, Kwang Nam Jin, Jae-Kwang Lim, Yun-Hyeon Kim, Ki Beom Kim,
Zicheng Jiang, Chuxiao Shao, Junqiang Lei, Shengqiang Zou, Hongqiu Pan,
Ye Gu, Guo Zhang, Jin Mo Goo, Soon Ho Yoon Preprint, Jul 24 2020
We retrospectively collected anonymized 176 chest CT scans of 131 reverse transcription-polymerase chain reaction (RT-PCR)-proven COVID-19 patients (mean age, 47.2±18.1 years; male to female ratio, 59%:41%) that were obtained at 13 Korean and Chinese institutions from 23 January to 15 March 2020.
Preparation of training CT data
The CT images were uploaded to a commercially available software program for semi-automatic segmentation (MEDIP PRO v2.0.0.0, MEDICALIP Co. Ltd., Seoul, Korea). The lung parenchyma was segmented by a previously developed deep neural network (DeepCatch v1.0.0.0, MEDICALIP Co. Ltd., Seoul, Korea; submitted), which automatically extracts lung parenchyma with an accuracy higher than 99% in CT images containing extensive lung disease. All parenchymal abnormalities of COVID-19 were initially segmented by two technicians. After reviewing the tentative lung and lesion masks, one of the two thoracic radiologists determined the presence of COVID-19 lesions and adjusted the masks in every axial CT image slice. The radiologists excluded parenchymal lesions other than COVID-19, such as peripheral reticulations and honeycombing, tuberculous sequelae, calcified nodules, dependent densities, pleural effusions, and areas of motion artifacts.
Development of the deep neural network
The 176 CT scans were randomly assigned into one of the three following data sets: 146 cases for the training set, 10 cases for the tuning set, and 20 cases for the internal validation set. A majority of the CT scans consisted of 1-mm-section CT images with standard- to low-dose CT protocols. Data were preprocessed by changing the Hounsfield unit (HU) values of the area outside the lung to -3024, and axial slices without the lung were not included in the training set. In total, 24,915 slices of axial data with areas of pneumonia and 30,711 slices of axial data without areas of pneumonia were available in the dataset. The training data were normalized by using the lung window setting.
Our 2D U-Net received an input with a size of 512×512×1 and consisted of initial convolutions, four encoders, four decoders, and a final convolution. Except for the final convolution, which was a 1×1 convolution, every convolutional layer consisted of a 3×3 convolution followed by batch normalization and the rectified linear unit (ReLu) activation function. For decoders, upsampling with bilinear interpolation was used, followed by concatenation to conserve information before down-sampling.
The Kaiming He initialization method was used for weight initialization. A sigmoid function was used in the final layer, and the model was trained using the stochastic gradient descent algorithm and the binary cross entropy loss function. After the training was completed, the tuning dataset was used to choose the best weight, which was saved after each epoch.
The 2D U-Net was distributed as free standalone software (MEDIP
COVID19) on 18 March 2020 and updated with the current version of
v1.2.0.0 on 27 April. The software conducted a quantification of
pneumonia in 1 minute with the recommended speculations. Based on the area of COVID-19 pneumonia segmented by the network,
the software measured the volume of the pneumonia and lung parenchyma
and automatically calculated the extent (%) and weight (g) of pneumonia.
0 件のコメント:
コメントを投稿