Explainable AI Model for COVID-19 Diagnosis Through Joint Deep Learning and Radiomics
Presentations
MO-IePD-TRACK 4-3 (Monday, 7/26/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: Deep learning based computer-aided diagnosis is a promising and efficient tool for the early detection of the Coronavirus Disease 2019 (COVID-19). However, the “black box” characteristic makes it difficult for clinicians to comprehend. This study aims to develop an explainable automatic COVID-19 detection system by combining deep learning classification and radiomics feature analysis.
Methods: A dataset of 4671 Chest X-rays from multiple public datasets was used in this study. A deep learning model was developed by inserting deformable convolutional layer into the last denseblock in DenseNet121. During training, data augmentation was applied to training data to prevent overfitting. The proposed deformable convolutional neural network predicted the category of the input Chest X-ray and output a series of visualization results with Grad-CAM++ method. A mask for the region of interest (ROI) was generated via element-wise multiplication between lung region mask and visualized attention mask. Pyradiomics was then used to extract the features of the ROI. Correlation between each feature and category was calculated by using statistical methods. Ten features were selected and evaluated to describe the variance between categories based on statistical scores and random forest modeling.
Results: The proposed network obtained an overall test accuracy of 95.0%. The recall and precision of COVID-19 were both above 0.95. Meanwhile, saliency maps with bounding boxes were provided for interpretability. In radiomics feature analysis, all selected features had a p-value less than 0.01. The feature based classifier achieved 79.6% on testing set, which proved the validity and consistency of using selected features to represent different categories.
Conclusion: We have developed an explainable deep learning based COVID-19 Chest X-ray diagnosis method. To the best of the author’s knowledge, our method is the first of its kind to combine class activation map and radiomics features to achieve the interpretability of convolutional neural network.
Funding Support, Disclosures, and Conflict of Interest: HMRF COVID190211
ePosters
Keywords
X Rays, Computer Vision, Localization
Taxonomy
IM- X-Ray: Machine learning, computer vision
Contact Email