Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2020 Apr;8(7):450.
doi: 10.21037/atm.2020.03.132.

Deep Learning for Detecting Corona Virus Disease 2019 (COVID-19) on High-Resolution Computed Tomography: A Pilot Study

Affiliations
Free PMC article

Deep Learning for Detecting Corona Virus Disease 2019 (COVID-19) on High-Resolution Computed Tomography: A Pilot Study

Shuyi Yang et al. Ann Transl Med. .
Free PMC article

Abstract

Background: To evaluate the diagnostic efficacy of Densely Connected Convolutional Networks (DenseNet) for detection of COVID-19 features on high resolution computed tomography (HRCT).

Methods: The Ethic Committee of our institution approved the protocol of this study and waived the requirement for patient informed consent. Two hundreds and ninety-five patients were enrolled in this study (healthy person: 149; COVID-19 patients: 146), which were divided into three separate non-overlapping cohorts (training set, n=135, healthy person, n=69, patients, n=66; validation set, n=20, healthy person, n=10, patients, n=10; test set, n=140, healthy person, n=70, patients, n=70). The DenseNet was trained and tested to classify the images as having manifestation of COVID-19 or as healthy. A radiologist also blindly evaluated all the test images and rechecked the misdiagnosed cases by DenseNet. Receiver operating characteristic curves (ROC) and areas under the curve (AUCs) were used to assess the model performance. The sensitivity, specificity and accuracy of DenseNet model and radiologist were also calculated.

Results: The DenseNet algorithm model yielded an AUC of 0.99 (95% CI: 0.958-1.0) in the validation set and 0.98 (95% CI: 0.972-0.995) in the test set. The threshold value was selected as 0.8, while for validation and test sets, the accuracies were 95% and 92%, the sensitivities were 100% and 97%, the specificities were 90% and 87%, and the F1 values were 95% and 93%, respectively. The sensitivity of radiologist was 94%, the specificity was 96%, while the accuracy was 95%.

Conclusions: Deep learning (DL) with DenseNet can accurately classify COVID-19 on HRCT with an AUC of 0.98, which can reduce the miss diagnosis rate (combined with radiologists' evaluation) and radiologists' workload.

Keywords: COVID-19; deep learning (DL); high resolution computed tomography (HRCT).

Conflict of interest statement

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at http://dx.doi.org/10.21037/atm.2020.03.132). The authors have no conflicts of interest to declare.

Figures

Figure 1
Figure 1
The workflow of DenseNet with 4 blocks. Dense Block 1 illustrate a three-layer block with dense connectivity. Pooling and Linear refer to global average pooling and fully connected layer.
Figure 2
Figure 2
CAM calculated by output feature maps of the last convolutional layer. (A) HRCT shows GGOs with consolidation in the 3 segments of the lung (→); (B) The red and light blue regions represent areas activated by the DenseNet, while the dark purple background represents areas that are not activated. This shows that the DenseNet is focusing on parts of the image where the disease is present (→). HRCT, high resolution CT.
Figure 3
Figure 3
Receiver operating characteristic plots for COVID-19 identification for the DenseNet algorithm (A: validation set; B: test set).
Figure 4
Figure 4
The sensitivity, specificity, accuracy and F1 value along with the change of threshold value (validation set).

Similar articles

See all similar articles
Feedback