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

Background: Hypoxia is a potentially life-threatening condition that can be seen in pneumonia patients.

Objective: We aimed to develop and test 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 of 605 COVID-19 cases admitted to Al Ain Hospital from 24 February to 1 July 2020 into 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, $HC{O}_3^{-}$ , K ⁺, Na ⁺, anion gap, C-reactive protein) served as ground truth.

Results: Radiologic findings in the lungs of COVID-19 patients provide reliable assessments of functional status with clinical utility. 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 of the models trained on single-projection radiograms was approximately 11÷12% and it dropped by 0.5÷1% if both projections were 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 had 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 were 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). However, none of the differences between 3D and 2D were statistically significant.

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.