Deep learning attention-guided radiomics for COVID-19 chest radiograph classification

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate assessment of lung involvement in coronavirus disease 2019 (COVID-19) using chest radiographs (CXR) is crucial for patient management.
• Differentiating COVID-19 from non-COVID-19 pneumonia and normal CXRs presents a diagnostic challenge.

Purpose of the Study:

• To develop and evaluate a novel two-step feature merging method for integrating deep learning and radiomics features.
• To improve the accuracy of classifying COVID-19, non-COVID-19 pneumonia, and normal chest radiographs.

Main Methods:

• A deformable convolutional neural network (CNN) was employed as a feature extractor to obtain deep learning latent representation (DLR) features.
• Radiomics features were extracted from regions of interest identified by the deformable CNN's attention mechanism.
• DLR and radiomics features were concatenated to create a merged feature set for classification.

Main Results:

• The merged feature set achieved an overall average accuracy of 91.0% for three-class classification, a 0.6% improvement over DLR-only classification.
• Classification of COVID-19 yielded high recall (0.926) and precision (0.976).
• The feature merging method demonstrated a statistically significant improvement in classification performance (P < 0.0001).

Conclusions:

• A two-step framework effectively integrates DLR and radiomics features for COVID-19 classification using chest radiographs.
• The proposed feature merging strategy enhances classification performance compared to using deep learning features alone.