Purpose: To characterize COVID-19 by six presentations – ground glass opacity (GGO), consolidation, nodule, linear/reticular, crazy paving, and architectural distortion – on chest CT for COVID-19 patients.
Methods: Transfer learning exhibits robust performance when labeled images are limited. A transfer learning approach was implemented through feature extraction using the VGG-19 deep convolutional neural network (CNN) pretrained on the ImageNet database. Features were extracted from the pooling layers of the network on CT sections from 41 patients (221 scans) serving as input. Each patient was confirmed to have COVID-19 via PCR test. Multi-label classification methods of Binary Relevance (BR), Classifier Chains (CC), and Label Powersets (LP) with support vector machines were utilized to predict the multi-label output of COVID-19 presentation to distinguish if a given COVID-19 presentation was present or not. ROC analysis was used to evaluate the classifier performance.
Results: In the task of distinguishing COVID-19 on thoracic CT, the CC pipeline outperformed BR and LP classifications as determined by F1-score, Hamming loss, and relative AUC. Classification of GGO yielded an area under the ROC (AUC) of 0.95 ± 0.015, consolidation yielded an AUC of 0.60 ± 0.024, nodule yielded an AUC of 0.92 ± 0.024, consolidation linear/reticular an AUC of 0.64 ± 0.021, crazy paving yielded an AUC of 0.75 ± 0.025, and architectural distortion yielded an AUC of 0.76 ± 0.042.
Conclusion: The proposed deep learning algorithm with VGG-19 network can identify and differentiate COVID-19 patterns on CT images, potentially improving prognostic evaluations and informing treatment decision-making for COVID-19 patients. Future methods include improvements to the deep network feature extraction, such as fine-tuning and ensembled networks, as well as an investigation of the clinical implications of COVID-19 CT patterns on an additional 916 cases recently acquired (1935 imaging studies).