The role of imaging AI and CT in COVID-19

Apr 28, 2020

This post aims to elucidate the clinical role of icolung, the first CE marked AI software for the quantification of thorax CT scans. We discuss the role of CT imaging in COVID, the role of AI for the support of CT interpretation and the positioning of icolung.

The role of CT imaging in COVID-19

Currently, there is no consensus of the role of computed tomography (CT), as also reflected by the different positions of the radiological societies around the world. We discuss its potential role for diagnosis and for risk stratification.

Diagnosis

COVID-19 is typically diagnosed with reverse-transcription polymerase chain reaction (RT-PCR) testing. However, RT-PCR also has some important limitations. First of all, the test is not universally available. Furthermore turnaround times can be lengthy and reported sensitivities vary.

On the one hand, in the lung CT can detect certain characteristic manifestations associated with COVID-19. As such, studies have demonstrated that CT is more sensitive to detect COVID-19, with reported values of 97% (Ai et al. 2020)1 and 98% (Fang et al. 2020)9, compared to 71% for RT-PCR (Fang et al. 2020)9.

On the other hand, typical COVID associated abnormalities visible on CT are also present in numerous other pathologies, resulting in a lower specificity of CT compared to RT-PCR (which is highly specific) (Fang et al. 2020)9. Hence, if sufficient and rapid RT-PCR testing is available, initial diagnosis or screening with CT is currently not recommended by most radiological organizations.

However, within an environment where the reliability of COVID-19 testing is limited and turnaround times are long, imaging is indicated. Besides CT imaging, chest radiographs (CXR) could be considered as well, although it is estimated that CXR has little value in the first stages of the disease. Indeed, CXR may even present normal in early or mild disease (Wong et al. 202019, Ng et al. 202014), as illustrated in the example of below. Wong et al.19 report a sensitivity of 69% for baseline CXR. CT imaging is more sensitive to detect early pneumonic changes.

Figure 1. Comparison of chest radiograph (A) and CT thorax coronal image (B). The ground glass opacities in the right lower lobe periphery on the CT (red arrows) are not visible on the chest radiograph, which was taken 1 hour apart from the first study.

Risk stratification

Many societies, as well as the multinational consensus statement of the Fleischner society (Rubin et al., 2020)17 see a role for CT to assess the pulmonary status, especially in patients with moderate-to-severe respiratory disease.

The pulmonary status on CT is assessed by describing typical COVID associated patterns. The RSNA Expert Consensus Statement provides a guidance to radiologists in reporting CT findings potentially attributable to COVID-19 pneumonia, including standardized language (Simpson et al. 2020)18.

The degree of lung involvement on CT is an essential aspect of the pulmonary status as it correlates well with both severity of the disease (Radiology Assistent15, Zhang et al. 202020, Li, K. et al. 202012) as well as with disease outcome (Colombi et al. 20206, Revel et al. 202016). The degree of affected lung tissue can be expressed by scoring the percentages of each of the five lobes that are involved. Each lobe receives a score between 0 and 5 (Radiology Assistent15, Ding et al. 20207) or between 0 and 4 (Chung et al. 20205, Chaganti et al. 20202, Li, K. et al. 202012, Zhang et al. 202020), resulting in a total score on 25 or 20.

Hence, CT allows the reliable assessment of pulmonary status to facilitate risk stratification for clinical worsening. This stratification supports the triage of patients within the hospital.

Since CXR is less sensitive to pick up lung lesions, especially ground glass opacities, its use for assessing lung involvement is limited. To our knowledge, there are no publications yet showing a correlation between CXR findings and severity or outcome of the disease.

Monitoring

In case of clinical worsening, imaging is advised to assess for COVID-19 progression or secondary cardiopulmonary abnormalities such as pulmonary embolism, superimposed bacterial pneumonia, or heart failure that can potentially be secondary to COVID-19 myocardial injury (Rubin et al 2020)17.

The role of AI for CT imaging in COVID-19

Artificial intelligence (AI)-powered analysis of chest scans has the potential to alleviate the growing burden on radiologists, who must review and prioritize a rising number of patient chest scans each day. In the future, the technology might also help predicting which patients are most likely to need mechanical ventilation or medicines.

AI tools to help radiologists are being developed by many initiatives. Today, we can distinguish AI that supports diagnosis and AI that assesses the severity of the disease to support risk stratification. Here, we will discuss the AI algorithms in order of increasing level of annotation intensity during training. The more detailed the annotations, the more context is given to the algorithm.

Diagnostic support: COVID-19 detection

Deep learning-based COVID detection networks start from lung CT images, often after lung extraction, in order to classify the images based on a convolutional neural network. Making use of RestNet50 (He et al. 201611), Li, L. et al.13 developed COVNet to distinguish between COVID, community acquired pneumonia and non-pneumonias with an architecture shown in Fig. 2.

Figure 2. Architecture of COVNet to detect COVID based on thorax CT scan (Li, L. et al. 2020)13.

Over 4000 images were used to train the COVNet, with each image individually labeled ("weak labels"). An AUC of 0.96, sensitivity of 90% and specificity of 96% on about 400 images was reported.

In order to train a 2D network, called ResNet-50, individual slices potentially containing COVID associated tissue were labeled as normal (n=1036) or abnormal (n=829) (Gozes et al. 2020)10. An AUC of 0.996 (sensitivity of 98% and specificity of 92%) on 157 images was reported. In order to produce visual explanation for network decisions, the Grad-cam technique is used, allowing "heat maps of COVID associated tissue".

Figure 3. Visual representation of the decision made in by the network of Gozes et al10.

To train their 2D Unet++ network, Chen et al.4 performed a more labour-intense labeling by drawing a bounding box around COVID associated lesions. Good results on a relative small training dataset of 691 slices of COVID-19 pneumonia and 300 slices of normal controls were achieved. Chen et al.4 reported a sensitivity of 100% and a specificity of 94% for detecting COVID.

Several companies offer COVID-19 detection solutions, e.g RADLogics, Alibaba, DetectED. We are not aware of the regulatory status of these solutions.

An important limitation of multiple COVID detection algorithms is that it is uncertain whether they can properly distinguish COVID from other lung disorders with similar abnormality patterns. Indeed, COVID associated abnormalities, such are ground glass opacities and consolidations, are not characteristic for only COVID. Therefore, training and test data should include sufficient cases of non-COVID lung disorders.

Li, L. et al. 202013 included community-acquired pneumonia and non-pneumonia cases, resulting in seemingly lower performance. However - given the amount of images the network was trained on together with the quality of their validation technique - we estimate their network as clinically more relevant to support diagnosis.

Risk assessment: quantification and classification

To allow the assessment of risks for clinical worsening, the amount of affected tissue is to be quantified as well. Therefore, images need to be segmented in COVID associated abnormalities and background. AI algorithms are able to automatically segment images when trained on manually delineated images. This manual delineation of COVID associated abnormalities is very labour-intense.

Chaganti et al.2 proposed a DenseUNet to segment the COVID associated abnormalities, as illustrated in Fig. 4.

Figure 4: Overview of the deep learning based segmentation algorithm (Chaganti et al. 2020)2.

The algorithm was trained on 613 manually delineated CT scans (160 COVID-19, 172 viral pneumonia, and 296 ILD). The reported Pearson Correlation Coefficient between method prediction and ground truth is 0.98 for the total abnormalities.

Chassagnon et al.3 present CovidENet as a combination of a 2D slice-based and 3D patch-based ensemble architectures, trained on 23423 slices. They reported that CovidENet performed equally well as trained radiologists, with a Dice coefficient of 0.7.

Several initiatives are working on other quantification tools for lung CTs. Some companies, like Infervision and HY, have lung quantification software available. It is unclear whether this software was trained on COVID-19 cases.

Quantification of lung involvement with icolung

icolung is the first, and to the best of our knowledge the only, CE marked AI-based quantification tool. It quantifies the abnormalities found in COVID-19, and represents the results in an easy-to-interpret quantitative report.

Figure 5: Quantitative report as an output of icolung 0.2.

This report is seamlessly integrated into the clinical workflow, as non-contrast thorax CT data (parenchymal window) are sent from the PACS or CT scanner to the application. Subsequently, quantitative results and annotated images are sent back into the PACS before the radiologist starts the reading.

The algorithm is iteratively updated, when more training data is available. For example, icolung 0.2 (available since April 23rd) is trained on a relative small training dataset of 691 slices of COVID-19 pneumonia and 300 slices of normal controls (Chen et al. 2020)4; icolung 0.3 (available soon) is trained on more than 40 000 slices. For the performance of icolung, we refer to the technical documentation.

Clinical benefits of icolung

Since icolung is a CE marked medical device, it can be used in clinical routine practice in the EEA, and, hence, allows the fast and reliable quantification of lung involvement with possible impact on clinical decisions.

The results of icolung (quantitative reports & annotated images) are available for the radiologist. As a consequence, radiologists can report a more complete pulmonary status, including lung involvement (expressed as percentages) and, possibly, the severity score. This lung involvement correlates with clinical status and contributes to outcome prediction (see "The role of CT imaging in COVID-19"). Using this additional information, pulmonologists and intensive care physicians are able to take more informed treatment decisions. From Colombi et al6, it is known that by taking this information into account, the specificity to predict negative outcome (ICU admission/death) increases. This allows clinicians to better identify patients who might be sent home or to a non-ICU bed. As a result, ICU beds could, potentially, be saved for patients who need them most (high lung involvement and severe clinical symptoms).

Moreover, it is expected that clinicians will also be able to better identify patients who might not need mechanical ventilation. With the recent concerns that some hospitals have been too quick to put COVID-19 patients on mechanical ventilators, this allows a better use of resources, with better patient outcomes (Dondorp et al 2020)8.

Conclusion

icolung is a valuable clinical tool for the assessment of lung involvement in COVID patients coming to the hospital. It allows a better risk-based stratification of these patients, and, potentially, allows to save ICU beds for those needing them most.

References

1. Ai et al., https://doi.org/10.1148/radiol.2020200642 (2020)

2. Chaganti et al. https://arxiv.org/pdf/2004.01279.pdf (2020)

3. Chassagnon et al. https://doi.org/10.1101/2020.04.17.20069187 (2020)

4. Chen et al., https://doi.org/10.1101/2020.02.25.20021568 (2020).

5. Chung et al.https://doi.org/10.1148/radiol.2020200230 (2020)

6. Colombi et al. https://doi.org/10.1148/radiol.2020201433 (2020)

7. Ding et al. https://doi.org/10.1016/j.ejrad.2020.109009 (2020)

8. Dondorp et al. https://doi.org/10.4269/ajtmh.20-0283 (2020)

9. Fang et al., https://doi.org/10.1148/radiol.2020200432 (2020)

10. Gozes et al. https://arxiv.org/abs/2003.05037 (2020)

11. He et al., https://doi.org/10.1109/CVPR.2016.90 (2016)

12. Li, K. et al. https://doi.org/10.1007/s00330-020-06817-6 (2020)

13. Li, L. et al. https://doi.org/10.1148/radiol.2020200905 (2020)

14. Ng et al. https://doi.org/10.1148/ryct.2020200034 (2020)

15. Radiology Assistent https://radiologyassistant.nl/chest/lk-jg-1 (accessed April 26th, 2020)

16. Revel et al. https://doi.org/10.1007/s00330-020-06865-y (2020)

17. Rubin et al. https://doi.org/10.1148/radiol.2020201365 (2020)

18. Simpson et al. https://doi.org/10.1148/ryct.2020200152 (2020)

19. Wong et al. https://doi.org/10.1148/radiol.2020201160 (2020)

20. Zhang et al. https://doi.org/10.1007/s00330-020-06854-1 (2020)