AUTHOR=Statsenko Yauhen , Habuza Tetiana , Talako Tatsiana , Pazniak Mikalai , Likhorad Elena , Pazniak Aleh , Beliakouski Pavel , Gelovani Juri G. , Gorkom Klaus Neidl-Van , Almansoori Taleb M. , Al Zahmi Fatmah , Qandil Dana Sharif , Zaki Nazar , Elyassami Sanaa , Ponomareva Anna , Loney Tom , Naidoo Nerissa , Mannaerts Guido Hein Huib , Al Koteesh Jamal , Ljubisavljevic Milos R. , Das Karuna M. 

TITLE=Deep Learning-Based Automatic Assessment of Lung Impairment in COVID-19 Pneumonia: Predicting Markers of Hypoxia With Computer Vision

JOURNAL=Frontiers in Medicine

VOLUME=Volume 9 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2022.882190

DOI=10.3389/fmed.2022.882190

ISSN=2296-858X

ABSTRACT=Background: Hypoxia is a potentially life-threatening condition that can be seen in pneumonia patients.
Objective: We aimed to work out an automatic assessment of lung impairment in COVID- 19 associated pneumonia with machine learning regression models that predict markers of respiratory and cardiovascular functioning from radiograms and lung CT.
Materials and methods: We enrolled a total number of 605 cases admitted to Al Ain Hospital from 24 February to 1 July 2020 for the study. The inclusion criteria were as follows: age 18 years; inpatient admission; PCR positive for SARS-CoV-2; lung CT available at PACS. We designed a CNN-based regression model to predict systemic oxygenation markers from lung CT and 2D diagnostic images of the chest. The 2D images generated by averaging CT scans were
analogous to the frontal and lateral view radiograms. The functional (heart and breath rate, blood pressure) and biochemical findings (SpO2, HCO3−, K+, Na+, anion gap, C-reactive protein) served as ground truth.
Results: Radiologic findings in the lungs of COVID-19 patients assess functional status reliably. If fed to ML models, the sagittal view radiograms reflect dyspnea more accurately than the coronal view radiograms due to the smaller size and the lower model complexity. Mean absolute error (MAE) of the models trained on single-projection radiograms is around 11 12% and it drops by 0.5   1% if both projections are used (11.97   9.23 vs 11.43   7.51%; p=0.70). Thus, the ML regression models based on 2D images acquired in multiple planes have slightly better performance. The data blending approach was as efficient as the voting regression technique: 10.90±6.72 vs 11.96±8.30%, p=0.94. The models trained on 3D images are more accurate than those on 2D: 8.27±4.13 and 11.75±8.26%, p=0.14 before lung extraction; 10.66±5.83 and 7.94±4.13%, p=0.18 after the extraction. The lung extraction boosts 3D model performance unsubstantially (from 8.27±4.13 to 7.94±4.13%; p=0.82).
Conclusion: The constructed ML algorithms can serve as models of structure-function association and pathophysiologic changes in COVID-19. The algorithms can improve risk evaluation and disease management especially after oxygen therapy that changes functional findings. Thus, the structural assessment of acute lung injury speaks of disease severity.