COVID-19 Detection Using Deep Learning
A walk-through of state-of-the-art in COVID-19 Detection
The outbreak of the COVID-19 pandemic has crippled the world’s healthcare systems. Although researchers have developed vaccines, it will take a fair amount of time to vaccinate each citizen. Moreover, with the emergence of new variations of the coronavirus, early detection and prevention of the spread is of utmost importance.
The current gold-standard test for detecting COVID-19 is the Reverse Transcriptase Polymerase Chain Reaction (RT-PCR). However, it takes about 6–8 hours to produce results while also being less sensitive [1].
“Saliva sensitivity was highest in samples collected during the first week of infection at 71.2% (95% CI, 62.6%- 78.8%) but decreased each subsequent week.”
— Source: [1]
Although the rapid antigen test can produce results within 15 minutes by detecting the IgG and IgM antibodies simultaneously in the human blood, it might take several days for the human body to form these antibodies. Thus, there is a risk of spreading the virus before being detected. Hence, as an alternative, an automated diagnosis tool is required that is sensitive and specific to the COVID-19 disease, producing fast predictions.
Machine Learning methods that work with X-Ray or CT-Scan images of the lungs have proven successful in terms of accuracy, sensitivity, time complexity, and reliability for detecting the onset of COVID-19 [2] while also being more accessible than the RT-PCR/Rapid Antigen-Antibody tests. Although a radiological test needs to be performed, arranging for such a test is easier than the RT-PCR test due to shortages.
Fang et al. [3] found the sensitivity of chest CT-Scans to be 98% compared to the 71% sensitivity of RT-PCR for COVID-19 detection.
Let us now discuss a few computer-aided diagnosis methods proposed in the literature for efficient COVID-19 detection.
COVID-19 Image Classification
The COVID-19 infection is distinctly visible as “ground-glass opacities (GGOs)” in a chest CT-scan image [12]. Two examples of CT-scan images (a COVID-case vs. a Non-COVID case) taken from an open-source dataset [4] is shown below.
The marked portion in the CT scan of the “COVID +ve” case shows the GGO, which is the discriminating feature of COVID-19. They represent tiny air sacs or alveoli filled with fluid and turning a shade of grey in the CT-scan, turning into a white consolidation in more severe cases [12].
A deep learning model trained to classify COVID and Non-COVID cases will focus on whether or not the CT-Scan contains these white consolidations.
Kundu et al. [5] employed an Ensemble Learning-based framework (Read more on Ensemble Learning here) to detect COVID-19 from CT-Scan images via image classification. The ensemble framework is based on a fuzzy learning algorithm, that intelligibly aggregates confidence scores of contributing models to generate a final prediction (Codes available on GitHub).
The authors employed four popularly used pre-trained models in [5] (which are VGG-11, GoogLeNet, SqueezeNet v1.1, and Wide ResNet-50–2) to generate confidence scores of predictions. They generated the heat maps for a few example images from these four pre-trained models to show that they localize the correct part of the CT-scan images. These images are shown below.
From the examples shown above, we see that regardless of which model is being used, the focus of the model is always on the two lobes of the lungs in the CT-scan images, where the GGOs occur in COVID infected patients [images (a) and (f) ]. In cases where the GGOs are absent in either lung lobes [images (k) and (p)], the models predict “Non-COVID” or “COVID -ve.”
The results obtained by the ensemble framework the authors proposed in [5] are shown in the table above. Consistent results were obtained in both “COVID” and “Non-COVID” cases, making the deep learning framework a reliable detector of COVID-19.
The results obtained by the different pre-trained models and the ensemble method proposed in [5] are shown in the table above. Notice that very high results (90%+) are obtained by each model, which is much higher than what is obtained from RT-PCR tests. This indicates the reliability of Machine Learning in a complex disease detection task.
Other classification-based COVID-19 detection frameworks include works like Chattopadhyay et al. [6], who used deep learning coupled with evolutionary algorithms; Garain et al. [7] who used a spiking neural network; and Mahmud et al. [8] who devised an end-to-end Convolutional Neural Network model for COVID detection for COVID-19 detection.
COVID-19 Disease Localization
Classifying a CT image into COVID and Non-COVID cases is often not intuitive for medical experts. This is because even the deep learning models are not perfect. They cannot classify every image perfectly, leading to significant adverse effects. Thus, clinicians prefer a disease localization map from the radiography images, which can aid them in their diagnosis and also help in judging the severity of the disease.
Want to get the most up-to-date news on all things Deep Learning? Subscribe to Deep Learning Weekly and get the latest news delivered right to your inbox.
Thus, several researchers have devised automated disease localization frameworks as well. The basic structure of such frameworks uses an image-to-image mapping model using an encoder-decoder pipeline for the semantic segmentation of the input image. Semantic segmentation refers to the pixel-level classification of an image, i.e., each pixel is determined to be of class “object” or “background” (in the case of binary segmentation).
One such work is the chest CT segmentation framework devised by Voulodimos et al. [9], where the authors investigate the efficacy of deep learning-based segmentation models for localizing GGOs in CT-Scan images. The examples below show the results.
The “Ground truth” images are the annotations made by medical experts. The localization outputs obtained by the different models (FCN-8s and U-Net) are satisfactorily close to the annotations made by the experts, justifying the reliability of automated methods. The semantic segmentation results, highlighting the GGOs in the chest CT-Scan images, obtained by the authors on some more example images are shown below.
Another work, by Fan et al. [10], takes this a step further to develop a multi-class segmentation framework, named “Inf-Net,” based on attention modules. An example of the result obtained by them is shown below. The red-colored markings show the GGOs formed in the lungs, and the green markings show the pulmonary consolidations, i.e., a late stage of the GGOs where the consolidated parts of the lungs are damaged.
Hybrid Segmentation-Classification Frameworks
Driven by the need for localization of the disease in radiography images to give medical experts a comprehensive outlook, and for the coarse classification of the images into “COVID” and “Non-COVID” cases to make the automated frameworks accessible even to non-experts, researchers have developed hybrid systems that perform both segmentation and classification of an input image. This enables the fast detection of COVID-19 using the classification set-up while also allowing experts to have a closer look at the CT-Scan in cases of ambiguity through the segmentation set-up.
Wang et al. [11] developed one such framework, based on weakly supervised learning, i.e., here, the authors had the classification labels available, but the segmentation ground truths were unavailable. The framework proposed in [11] is shown below.
And the segmentation and classification results obtained by the authors in [11] are shown below.
Segmentation Results:
Classification Results:
As can be seen from the results obtained above, both the segmentation and classification performances are remarkable. The authors have also made the codes for their implementation available via GitHub.
Conclusion
The primary concern of the COVID-19 disease is its fast-spreading nature, which needs to be curbed. Although the RT-PCR test is used as a gold-standard measure for COVID infection, the time it takes to produce results makes it difficult to conduct population-wide screening.
Thus, automated frameworks based on Machine Learning models are being developed by researchers to aid clinicians in the fast and accurate detection of COVID-19. Such models perform COVID detection with much higher sensitivity than the RT-PCR tests and in an exponentially shorter amount of time. The disease localization frameworks developed further give experts a comprehensive overview of the spread of the disease in the patient’s lungs. This aids clinicians in cases of ambiguous predictions by deep learning-based classification models.
References
[1] Congrave-Wilson, Zion, et al. “Change in Saliva RT-PCR Sensitivity Over the Course of SARS-CoV-2 Infection.” JAMA 326.11 (2021): 1065–1067.
[2] Kundu, Rohit, et al. “ET-NET: an ensemble of transfer learning models for prediction of COVID-19 infection through chest CT-scan images.” Multimedia Tools and Applications (2021): 1–20.
[3] Fang, Yicheng, et al. “Sensitivity of chest CT for COVID-19: comparison to RT-PCR.” Radiology 296.2 (2020): E115-E117.
[4] Angelov, Plamen, and Eduardo Almeida Soares. “SARS-CoV-2 CT-scan dataset: A large dataset of real patients CT scans for SARS-CoV-2 identification.” MedRxiv (2020).
[5] Kundu, Rohit, et al. “COVID-19 detection from lung CT-Scans using a fuzzy integral-based CNN ensemble.” Computers in Biology and Medicine 138 (2021): 104895.
[6] Chattopadhyay, Soham, et al. “COVID-19 detection by optimizing deep residual features with improved clustering-based golden ratio optimizer.” Diagnostics 11.2 (2021): 315.
[7] Garain, Avishek, et al. “Detection of COVID-19 from CT scan images: A spiking neural network-based approach.” Neural Computing and Applications (2021): 1–14.
[8] Mahmud, Tanvir, Md Awsafur Rahman, and Shaikh Anowarul Fattah. “CovXNet: A multi-dilation convolutional neural network for automatic COVID-19 and other pneumonia detection from chest X-ray images with transferable multi-receptive feature optimization.” Computers in biology and medicine 122 (2020): 103869.
[9] Voulodimos, Athanasios, et al. “Deep learning models for COVID-19 infected area segmentation in CT images.” The 14th PErvasive Technologies Related to Assistive Environments Conference. 2021.
[10] Fan, Deng-Ping, et al. “Inf-net: Automatic covid-19 lung infection segmentation from ct images.” IEEE Transactions on Medical Imaging 39.8 (2020): 2626–2637.
[11] Wang, Xinggang, et al. “A weakly-supervised framework for COVID-19 classification and lesion localization from chest CT.” IEEE transactions on medical imaging 39.8 (2020): 2615–2625.
[12] Schmitt, W., and E. Marchiori. “Covid-19: round and oval areas of ground-glass opacity.” Pulmonology 26.4 (2020): 246.
Editor’s Note: Heartbeat is a contributor-driven online publication and community dedicated to providing premier educational resources for data science, machine learning, and deep learning practitioners. We’re committed to supporting and inspiring developers and engineers from all walks of life.
Editorially independent, Heartbeat is sponsored and published by Comet, an MLOps platform that enables data scientists & ML teams to track, compare, explain, & optimize their experiments. We pay our contributors, and we don’t sell ads.
If you’d like to contribute, head on over to our call for contributors. You can also sign up to receive our weekly newsletters (Deep Learning Weekly and the Comet Newsletter), join us on Slack, and follow Comet on Twitter and LinkedIn for resources, events, and much more that will help you build better ML models, faster.
