5.1. CT image Dataset
The institutional review board of Ruian people’s hospital in Zhejiang Providence, China approved this retrospective study, and the written informed consent requirement was waived. From January 2020 to March 2020, 1611 CT slice images from 175 patients, including 61 COVID-19 patients, 27 CAP patients, and 87 normal control patients, were enrolled in this study (age ranges from 5 to 72 years). Figure 4 presents patient inclusion and exclusion criterion. Ultimately, 491 COVID-19, 138 CAP, and 982 normal control CT images were included. Figure 5 shows the typical examples of COVID-19, CAP, and Normal CT images.
Three senior radiologists retrospectively reviewed both COVID-19, CAP, and normal control patients who had undergone chest CT scans. Two radiologists independently assessed the images. If their diagnostic results differ from each other, the other radiologist would be invited to review the pulmonary CT images with controversial results and make the final decision by negotiation. All the CT images were generated from the Siemens Sensation 16-layer spiral CT (Siemens, Erlangen, Germany). The image format was Digital Imaging and Communications in Medicine (DICOM). The scan parameters were tube voltage 120 kV; tube current automatic regulation; 1-2 mm cross-sectional thickness; 1–2 mm cross-sectional distance; scan pitch 1.3; and 16 × 0.625 mm collimation.
5.2. Image Preprocessing
It is observed that heavy data imbalance exists in the collected datasets while imbalanced data cannot reflect the true distribution of three categories. It will directly affect the performance of image classification. Thus, data augmentation was performed. We rotated CAP images by 90°, 180°, and 270° and COVID-19 images by 90°. After that, we did mirror transformation to the original and rotated CAP images. Finally, the number of CAP images was expanded to 1104 and the number of COVID-19 images was expanded to 982. In conclusion, 3068 CT images were enrolled for training, validation, and testing (dataset available at https://ieee-dataport.org/documents/covid-19dataset).
To extract pulmonary parenchyma of chest CT images, an active contour model and Bayesian approach (available at https://github.com/HeewonChung92/LungSegmentation) was applied to automatically segment the lungs[34]. However, the lung segmentation method failed to segment ground-glass opacities (GGOs) and consolidation shadows near the lung boundary. Consequently, manual segmentation was performed on the inaccurately segmented images by radiologists. Figure 6 shows the examples of wrongly segmented CT images by active contour model and Bayesian approach and correctly segmented CT image with the help of radiologists. It is noteworthy that wrongly segmented CT images were utilized to evaluate the value of ground-glass opacity and consolidation shadows for distinguishing COVID-19 from CAP.
5.3. The proposed model design
A standard CNN has several convolution layers to capture image features, each of which was followed by a pooling layer to reduce the number of parameters, a fully connected layer for classification. By adding more layers, CNN can achieve better performance. For example, ResNet-101 yielded higher accuracy than ResNet-50[35] because deeper CNN networks can extract higher-level features. Similarly, more fine-grained features can be captured by wider CNN and more fine-grained patterns tend to be extracted with higher resolution images input to CNN. Accordingly, scaling up one of the three dimensions, including depth, width, and image resolution, is the most common way to improve classification accuracy.
However, the accuracy of the model gained rapid saturate by scaling only one of the three dimensions. To tackle this problem, Mingxing Tan and Quoc V. Le proposed EfficientNet[36], which uniformly scales all three dimensions of the CNN networks. Furthermore, EfficentNets achieve state-of-the-art accuracy in image classification with fewer model parameters and high efficiency.
Instead of initiating a deep model development from scratch, a more rational approach is to construct a deep learning network by fine-tuning already proven models. Therefore, EfficientNet-B1 was applied for the classification task, which consists of 16 mobile inverted bottleneck convolutions (MBConv), 2 convolutional (Conv) layers, 1 global average-pooling layer, and 1 fully connected layer(code available at https://codeocean.com/capsule/7306942/tree). For binary classification, the model ends with a classification layer with sigmoid activation to make predictions of COVID-19 versus CAP, while for three classed classification, the activation function was replaced with softmax.
Dropout with the probability of 0.2 was applied to the fully connected layer. In the process of training, data augmentation was performed dynamically again, including scaling, blurring, random brightness, and contrast manipulations. The training was performed for 30 epochs with a batch size of 64. The Adam optimizer with default parameters was utilized for weight updates and a cross-entropy loss function was implemented. The initial value of the learning rate is set to 0.0001, and it will immediately decrease by a factor of 10 once learning stagnates. The validation set is randomly split from the training set at a scale of 0.2 and validation was performed on every epoch. The ratio of the training dataset to the test dataset is 8:2. The checkpoint with the highest validation accuracy was retained as the optimal model for test set evaluation.
The proposed model has 16 MBConv layers, each of which is composed of Depthwise Separable Convolution and SENet [37]. Figure 7 demonstrates the architecture of MBConv. As is shown in Figure 4, DepthwiseConv is followed by batch normalization and Swish operations. Batch normalization operation is applied to accelerate the training procedure and increase the stability of the model. Swish is a self-gated activation function, which achieves better performance than rectified linear unit (ReLU) does on deeper models[38] . The calculation of Swish is given in equation (1):
![](data:image/png;base64,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)
The calculation of Swish is given in equation
Finally, the developed EfficientNet-B1 model consists of 6,637,953 parameters for binary classification and 6,763,395 parameters for 3-class classification. An illustration of the EfficientNet-B1 model implemented for classification is shown in Figure 8
5.4. Experiment design
We conducted experiments on two different CT image data to detect and classify COVID-19. First, the proposed EfficientNet-B1 was trained and evaluated for binary classification (COVID-19 and CAP). Simultaneously,another same deep learning model was trained to classify CT images into three classes: COVID-19, CAP, and Normal. For each experiment, CT images that were only segmented based on the active contour model and Bayesian approach were taken as the control group while the properly segmented images were leveraged as the experimental group.
5.5. Performance Evaluation
The model performance was evaluated based on the test set. The proposed method was assessed in terms of accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV), and F1-score. The calculations of evaluation indexes are given in equations (2)-(7):
![](data:image/png;base64,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)
where TP represents correctly classified positive class; TN denotes correctly classified negative class; FP means inaccurately classified positive class; FN indicates erroneously classified negative class.