Classification of COVID-19 in chest X-ray images using DeTraC deep convolutional neural network (2003.13815v3)

Abstract: Chest X-ray is the first imaging technique that plays an important role in the diagnosis of COVID-19 disease. Due to the high availability of large-scale annotated image datasets, great success has been achieved using convolutional neural networks (CNNs) for image recognition and classification. However, due to the limited availability of annotated medical images, the classification of medical images remains the biggest challenge in medical diagnosis. Thanks to transfer learning, an effective mechanism that can provide a promising solution by transferring knowledge from generic object recognition tasks to domain-specific tasks. In this paper, we validate and adapt our previously developed CNN, called Decompose, Transfer, and Compose (DeTraC), for the classification of COVID-19 chest X-ray images. DeTraC can deal with any irregularities in the image dataset by investigating its class boundaries using a class decomposition mechanism. The experimental results showed the capability of DeTraC in the detection of COVID-19 cases from a comprehensive image dataset collected from several hospitals around the world. High accuracy of 95.12% (with a sensitivity of 97.91%, a specificity of 91.87%, and a precision of 93.36%) was achieved by DeTraC in the detection of COVID-19 X-ray images from normal, and severe acute respiratory syndrome cases.

• The paper introduces an adapted DeTraC deep CNN that leverages class decomposition to refine decision boundaries for COVID-19 detection.
• The paper employs transfer learning from ResNet18 combined with k-means clustering to enhance accuracy, sensitivity, and specificity.
• The paper demonstrates that the enhanced model outperforms the standard ResNet18, supporting rapid and accurate COVID-19 diagnosis in clinical settings.

Classification of COVID-19 in Chest X-ray Images Using DeTraC Deep Convolutional Neural Network

The paper presents an adaptation and validation of the DeTraC (Decompose, Transfer, and Compose) deep convolutional neural network (CNN) architecture for the classification of COVID-19 from chest X-ray (CXR) images. The proposed solution addresses two critical challenges in medical imaging: the limited availability of annotated datasets and data irregularities, particularly in the context of COVID-19 detection.

Methodology

The DeTraC model comprises three primary phases:

1. Feature Extraction and Class Decomposition: The initial phase involves training a pre-trained CNN, in this case, ResNet18, in a shallow-tuning mode to extract deep local features from each image. To manage data irregularities, the framework employs a class decomposition approach using k-means clustering, subdividing each class in the dataset into k homogeneous subclasses. This helps in refining the decision boundaries of the classifier.
2. Transfer Learning: The second phase leverages transfer learning from the ImageNet pre-trained ResNet18. By freezing low-level layers and updating high-level layers, this method preserves generalizable features while adapting to the new task. For parameter optimization, a mini-batch stochastic gradient descent (mSGD) with cross-entropy loss function is employed.
3. Class Composition and Evaluation: The final phase involves re-composing the predictions from subclass labels back to the original class labels to provide the final classification output. Performance is assessed using accuracy, sensitivity, and specificity metrics.

Experimental Setup and Results

The authors utilized a combination of two datasets: 80 normal CXR images from the Japanese Society of Radiological Technology (JSRT) and 116 images (105 COVID-19 and 11 SARS) from Cohen's collection. These datasets were augmented and preprocessed to enhance the robustness of the model.

Applying the DeTraC-ResNet18 configuration, the model achieved impressive performance metrics:

• Accuracy: 95.12%
• Sensitivity: 97.91%
• Specificity: 91.87%

The results are visually supported by learning curves for both accuracy and loss, indicating high effectiveness and robustness.

Comparative Analysis

The paper contrasted DeTraC-ResNet18 with a vanilla ResNet18 model. The standalone ResNet18 network achieved an accuracy of 92.5%, with a notably lower sensitivity of 65.01% and a specificity of 94.3%. This comparison underscores the advantage of integrating class decomposition to manage data irregularities efficiently.

Theoretical and Practical Implications

The DeTraC model exemplifies how class decomposition can enhance CNN performance in the context of limited and irregular medical image datasets. By breaking down classes into more granular sub-classes, the model can better learn the nuanced structures within each class, leading to improved generalization and reduced misclassification.

In practical terms, deploying this model in clinical settings can aid radiologists in the rapid and accurate diagnosis of COVID-19 from chest X-rays, which is crucial for timely intervention and controlling the spread of the virus.

Future Directions

Several avenues for future work are recommended:

• Larger Datasets: Extending the validation with progressively larger and more diverse datasets can reinforce the model's robustness and reliability.
• Explainability: Integrating explainability methods will enhance the usability of the model by providing clinical professionals with clear insights into the decision-making process.
• Model Efficiency: Techniques such as model pruning and quantization can be employed to increase the computational efficiency, facilitating deployment on handheld or edge devices.

Conclusion

The paper effectively demonstrates the capability of the DeTraC deep CNN architecture for COVID-19 detection from CXR images. By addressing the inherent challenges of limited data availability and data irregularities, DeTraC has shown promising results, marking a significant contribution to the field of medical image classification. The integration of class decomposition is particularly notable for its ability to refine the decision boundaries, thereby enhancing the accuracy and robustness of the classification.

This work paves the way for further research and application in the domain, potentially contributing significantly to rapid and efficient COVID-19 diagnostics.