Automatic classification between COVID-19 pneumonia, non-COVID-19 pneumonia, and the healthy on chest X-ray image: combination of data augmentation methods

Introduction

• CXR has generally lower sensitivity than chest CT, making accurate COVID-19 pneumonia diagnosis more challenging on CXR.The study frames CXR as a commonly used, cost-effective screening modality compared with CT.Source: Introduction · pp. 5–6
• The study aims to classify COVID-19 pneumonia, non-COVID-19 pneumonia, and healthy cases from CXR images using a CNN.The system is intended to remain accurate and robust with limited publicly available COVID-19 CXR training data.Source: Introduction · pp. 6–7
• The proposed approach uses transfer learning and combines conventional data augmentation with mixup, while also examining EfficientNet and RICAP.Conventional augmentation includes methods such as flipping, shifting, and rotation.Source: Introduction · pp. 6–7

Material and Methods

• The study used anonymized data collected from public datasets.Source: Material and Methods · p. 7
• Institutional review board approval was waived under the regulations of the authors’ country.Source: Material and Methods · p. 7
• Informed consent was not required.Source: Material and Methods · p. 7

Dataset

• Dataset I supplied COVID-19 and some non-COVID-19 pneumonia images, while dataset II supplied healthy and additional non-COVID-19 pneumonia images.The sources were a public COVID-19 image repository and the RSNA Pneumonia Detection Challenge dataset.Source: Dataset · pp. 7–8
• The final dataset contained 215 COVID-19, 533 non-COVID-19 pneumonia, and 500 healthy CXR images.The non-COVID-19 group combined 33 images from dataset I with 500 randomly selected images from dataset II; 500 healthy images were also selected from dataset II.Source: Dataset · p. 8
• The 1,248 images were split into 998 training, 125 validation, and 125 test images.Source: Dataset · p. 8

Deep learning model and data augmentation

• ImageNet-pretrained VGG16 weights were used for transfer learning, with 10 VGG16 layers frozen and added pooling, fully connected, and dropout layers.A three-unit fully connected layer was added after dropout for three-category classification.Source: Deep learning model and data… · p. 9
• The VGG16 input was 220 × 220 pixels, with a 416-unit first fully connected layer and dropout probability 0.1.Source: Deep learning model and data… · p. 9
• Training used RMSprop at a 1.0 × 10^-4 learning rate, batch size 8, softmax loss, and early stopping with patience 7.Training was set to 100 epochs, with early stopping monitored using validation loss.Source: Deep learning model and data… · pp. 9–10
• The proposed VGG16 model used conventional augmentation combined with mixup, selected through random search among combinations that also included RICAP.The conventional transformations included horizontal flipping and 85%–115% scaling and shear; mixup was set to 0.1.Source: Deep learning model and data… · p. 10

Comparison with other pre-trained models and ablation study

• The comparison included ResNet-50, MobileNet, DenseNet-121, and EfficientNet alongside VGG16.The other four models were transferred without freezing their trainable parameters; freezing degraded their performance, and random search selected hyperparameters and augmentation combinations.Source: Comparison with other pre-trained… · pp. 10–11
• EfficientNet’s best model was selected from variants B0–B7 by random search.The model comparison also used random search to find optimal hyperparameters and augmentation combinations.Source: Comparison with other pre-trained… · p. 11
• VGG16 ablations tested no augmentation, conventional augmentation only, mixup only, and conventional augmentation plus mixup without freezing.The configurations with no augmentation, conventional augmentation only, and mixup only used freezing.Source: Comparison with other pre-trained… · p. 11

Performance evaluation

• Five runs with different random seeds evaluated three-category accuracy on a 125-image test set.COVID-19 sensitivity was also calculated for the VGG16-based model.Source: Performance evaluation · p. 11

Results

• 83.7% mean accuracy was achieved by the proposed VGG16-based model, while each of the four other pre-trained models had mean accuracy below 80%.The mean COVID-19 sensitivity for VGG16 was 90.9%.Source: Results · p. 12
• The combined augmentation methods outperformed single-type or no-augmentation configurations, and layer freezing was effective in the VGG16-based model.The proposed combination of conventional augmentation and mixup was slightly better than conventional augmentation with RICAP; adding RICAP to conventional augmentation and mixup was not as good as the proposed method.Source: Results · pp. 12–13

Discussion

• More than 80% three-category accuracy and more than 90% COVID-19 sensitivity were reported for the proposed method.The paper concludes that an accurate CADx system for three-category classification can be built.Source: Discussion · p. 13
• Conventional augmentation combined with mixup was more effective than single-type or no augmentation, and was the most effective combination examined.Conventional augmentation plus RICAP was slightly inferior, while adding RICAP to the proposed combination was not as good.Source: Discussion · pp. 13–14
• VGG16 was the most accurate of the pre-trained models for three-category classification, despite other models scoring higher on ImageNet.The authors suggest that more complex models or models with more trainable parameters may have overfit the small dataset.Source: Discussion · p. 14
• The authors state that layer freezing was effective only in VGG16.They note VGG16 had a simpler architecture than the other models and suggest other networks’ skip connections may hinder layer freezing.Source: Discussion · p. 14
• The authors caution that combining pediatric non-COVID-19 pneumonia CXRs with adult COVID-19 CXRs may let CNNs distinguish age rather than disease, so their study included only adult CXR images.Source: Discussion · pp. 14–15
• The study used public datasets, and overfitting may have occurred in external validation.The authors say pre-trained CNN effectiveness with small datasets should be further investigated; this study used 1,248 CXRs.Source: Discussion · pp. 13–15
• The CADx system was not used by clinicians, so its clinical usefulness was not validated.Source: Discussion · p. 15

• The study concludes that its proposed method can build an accurate CADx system for three-category classification of COVID-19 pneumonia, non-COVID-19 pneumonia, and healthy cases.Source: Conclusion · p. 15
• Combining two data-augmentation methods was more useful than using a single method or none.Source: Conclusion · p. 15
• The authors plan to investigate system performance with clinical CXR images from all three categories.Source: Conclusion · p. 15

Online content

• The online-content text indicates that the study’s source code and dataset are available.Source: Online content · p. 17

Funding

• The funding disclosure names JSPS KAKENHI and grant number 19H03599.Source: Funding · p. 17

Conflicts of Interest

• The authors declare no conflict of interest and state that the funders had no role in the study’s design, data work, interpretation, writing, or publication decision.Source: Conflicts of Interest · p. 17

Figures

• Representative CXR images show COVID-19 pneumonia, non-COVID-19 pneumonia, and no pneumonia.Source: Figures · pp. 23–25
• The model-outline figure identifies VGG16, ResNet-50, MobileNet, DenseNet-121, and EfficientNet as the pre-trained models used.Source: Figures · p. 26

Tables

• Table 1 summarizes patient characteristics and CXR attributes, including collected age and sex information.Source: Table 1 · p. 27, Tables · p. 27
• Table 2 presents results for five pre-trained models and three-category classification, including evaluations involving hyperparameters and augmentation combinations.Source: Table 2 · p. 29, Tables · p. 29
• 83.2% (104/125) accuracy was reported for the representative three-category test-set confusion matrix.Source: Table 3 · p. 30, Tables · p. 30
• Table 4 evaluates data-augmentation methods and freezing trainable parameters in the VGG16-based model; its note reports cell values as means ± standard deviations across five trials.Source: Table 4 · p. 31, Tables · p. 31

Supplementary material

• Supplementary Table S1 concerns RICAP as an augmentation method, with cell values reported as means of five trials.Source: Table S1 · p. 32, Supplementary material · p. 32