Table 1.

Summary of the studies.

Author, Year	Study Design	Aim	Endpoints	AI Model	Results
Nishiyama, 2020 [9]	single center retrospective	prediction of prognosis	To evaluate the relationship between CT volume of well-aerated lung region and prognosis in ARDS patients.	An automated lung volumetry software of lung CT scan to identify lung region volumes by CT attenuation densities.	Well-aerated lung regions showed a positive correlation with 28-day survival. Survival outcome was better for percentage of well-aerated lung region/predicted total lung capacity ≥40% than <40%.
Gresser, 2021 [10]	single center retrospective	prediction of prognosis	To assess the potential of AI-based CT assessment and clinical score to predict the need for ECMO therapy in COVID-19 ARDS.	CT software provides segmentation of lung lobes providing a CT severity score.	AI-based assessment of lung involvement on CT scans at hospital admission and the SOFA scoring, especially if combined, can be used as risk stratification tools for subsequent ECMO requirement.
Hermann, 2021 [11]	multicenter retrospective	alveolar recruitment	To compare the accuracy in the computation of recruitability on CT scan between automatic lung segmentation performed by a properly trained neural network and manual segmentation in ARDS and COVID-19.	A DL algorithm to automatically segment ARDS injured lungs to calculate the lung recruitment.	The AI segmentation showed the same degree of inaccuracy of the manual segmentation. The recruitability measured with manual and AI segmentation had a bias of +0.3% and −0.5% expressed as change in well-aerated tissue fraction.
Kang, 2021 [12]	single center retrospective	differential diagnosis	To train a DL classifier model to differentiate between COVID-19 and bacterial pneumonia based on automatic segmentation of lung and lesion regions.	A DL model with deformable convolution neural network architecture trained to differentiate lesion patches of COVID-19 from those of bacterial pneumonia on CT scan.	DL lung CT scan analysis with constructed lesion clusters achieved an accuracy of 91.2% for classifying COVID-19 and bacterial pneumonia patients.
Lanza, 2020 [13]	single center retrospective	prediction of prognosis	To test quantitative CT analysis using a semi-automated method as an outcome predictor in terms of need for oxygen support or intubation in COVID-19.	Quantitative CT analysis with a semi-automated segmentation algorithm that divides lungs into not aerated, poorly aerated, normally aerated and hyperinflated.	The amount of compromised lung volume can predict the need for oxygenation support (between 6–23% of compromised lung) and intubation (above 23%) and is a significant risk factor for in-hospital death.
Liu, 2020 [14]	single center retrospective	prediction of prognosis	To quantify pneumonia lesions by CT (% of ground-glass, semi-consolidation and consolidation volume) in the early days to predict progression to severe illness using AI algorithms in COVID-19.	CT quantitative analysis combines a fully convolutional network with adopting thresholding and morphological operations for segmentation of lung and pneumonia lesions.	CT features on day 0 and 4, and their changes from day 0 to day 4, showed predictive capability for severe illness within a 28-day follow up. CT quantification of pneumonia lesions can early and non-invasively predict the progression to severe illness.
Pennati, 2023 [2]	single center retrospective	alveolar recruitment	To develop and validate classifier models to identify patients with a high percentage of potentially recruitable lung from readily available clinical data and from a single CT scan quantitative analysis at ICU admission.	Four ML algorithms (Logistic regression, Support Vector Machine, Random Forest, XGboost) to predict lung recruitment starting from a single CT scan obtained at 5 cm H2O at ICU admission.	The use of the four ML algorithms based on a CT scan at 5 cm H2O were able to classify lung recruiter patients with similar AUC as the ML algorithm, based on the combination of lung mechanics, gas exchange and CT data.
Penarrubia, 2023 [15]	single center retrospective	alveolar recruitment	To assess both intra- and inter-observer smallest real difference exceeding measurement error of recruitment using both human and ML on low-dose CT scans acquired at 5 and 15 cm H2O of PEEP in ARDS.	ML lung segmentation algorithm on CT scan to compute alveolar recruitment at 5 and 15 cm H2O of PEEP.	Human–machine and human–human inter-observer measurement errors were similar, suggesting that ML segmentation algorithms are valid alternative to humans for quantifying alveolar recruitment on CT.
Lopes, 2021 [16]	multicenter retrospective (study protocol)	prediction of prognosis	To develop a ML based on clinical, radiological and epidemiological data to predict the severity prognosis (ICU admission, intubation) in COVID-19.	A ML model receives a lung CT as input and outputs the stratification of lung parenchyma, discerning regions of the lungs with different densities.	Study in progress
Puhr-Westerheide, 2022 [17]	single center retrospective	diagnosis	To compare AI-based quantitative CT severity score to SOFA score in predicting in-hospital mortality at ICU admission in COVID-19 ARDS patients.	AI-based lung injury assessment on CT scan for the diagnostic performance to predict in-hospital mortality.	CT severity score was not associated to in-hospital mortality prediction, whereas the SOFA score showed a significant association.
Röhrich, 2021 [18]	single center prospective	prediction of prognosis	To develop a ML model for the early ARDS prediction from the first CT scan of trauma patients at hospital admission.	A ML model with convolutional neural network (radiomics) approach to automatically delineate the lung at lung CT to predict future ARDS.	The ML model with radiomics score resulted in a higher AUC (0.79) compared to injury severity score (0.66) and abbreviated injury score of the thorax (0.68) in prediction of ARDS. The radiomics score achieved a sensitivity and a specificity of 0.80 and 0.76.
Sarkar, 2023 [19]	single center retrospective	diagnosis and prediction of prognosis	To train and validate DL models to quantify pulmonary contusion as a percentage of total lung volume and assess the relationship between automated Lung Contusion Index and relevant clinical outcomes (ICU LoS and mechanical ventilation time).	DL model for automated CT scan segmentation to quantify the percent lung involvement indexed to total lung volumes.	Automated Lung Contusion Index was associated with ARDS, longer ICU LoS and longer mechanical ventilation time. Automated Lung Contusion Index and clinical variables predicted ARDS with an AUC of 0.70, while automated Lung Contusion Index alone predicted ARDS with an AUC of 0.68.
Wang, 2020 [20]	retrospective study	diagnosis	To explore the relationship between the quantitative analysis results and the ARDS existence, using an automatic quantitative analysis model based on DL segmentation model in COVID-19.	DL model to provide an automatic quantitative analysis of infection regions on lung CT to assess their density and location.	The total volume and density of the lung infectious regions were not related to ARDS. The proportion of lesion density was associated with increased risk of ARDS in COVID-19.
Zhang, 2020 [21]	single center retrospective	diagnosis	To compare the performance of the three DL models and determine which model is more diagnostic.	Three DL models (VGG, Resnet and EfficientNet) are used to classify LUS images of pneumonia according to different clinical stages based on a self-made image dataset.	EfficientNet showed to be the best model providing the best accuracy for 3 and 4 clinical stages of pneumonia, with an accuracy of 94.62% and 91.18%, respectively. The best classification accuracy of 8 clinical features of pneumonia at LUS images was 82.75%.
Baloescu, 2020 [22]	single center retrospective	diagnosis	To test the DL algorithm to quantify the assessment of B lines in LUS images from a database of patients presenting at ED with dyspnea or chest pain and to compare the algorithm to expert human interpretation.	A DL model is trained and developed based on a dataset of LUS clips to assess presence/absence of B lines and severity classification.	The accuracy in detecting B lines was 94% with a kappa of 0.88; the accuracy of the severity assessment was 56% with a kappa of 0.65.
Born, 2021 [23]	multicenter retrospective	differential diagnosis	To compare different AI models for the differential diagnosis of COVID-19 pneumonia and bacterial pneumonia.	Five AI models (VGG, VGG-CAM, NASNetMobile, VGG-segement, Segment-Enc) are tested on a dataset of LUS images and videos of healthy controls and patients affected by COVID-19 and bacterial pneumonia and compared in terms of recall, precision, specificity and F1 scores.	Two models (VGG and VGG-CAMI) had an accuracy of 88 ± 5% in distinguishing COVID-19 pneumonia and bacterial pneumonia.
Arntfield, 2021 [24]	multicenter retrospective	differential diagnosis	To compare the DL model and the surveyed LUS-competent physicians in the ability of discriminating pathological LUS imaging	A DL convolutional neural network model is trained on LUS images with B lines to discriminate between COVID-19 ARDS, non-COVID ARDS and hydrostatic pulmonary edema and compared with surveyed LUS-competent physicians.	The DL model showed an ability to discriminate between COVID-19 ARDS (AUC 1.0), non-COVID ARDS (AUC 0.934) and pulmonary edema (AUC 1.0) better than physician ability (AUCs 0.697, 0.704, 0.967).
Ebadi, 2021 [25]	multicenter retrospective	differential diagnosis	To compare the DL classifier model against ground truth classification provided by expert radiologists and clinicians.	A DL method based on the Kinetics-I3D network. classifies an entire LUS scan, without the use of pre-processing or a frame-by-frame analysis, for automatic detection of ARDS features present in pneumonia and COVID-19 patients (A lines, B lines, consolidation and pleural effusion).	The DL model showed an accuracy of 90% and a precision score of 95% with the use of 5-fold cross validation.

AI: artificial intelligence; ARDS: acute respiratory distress syndrome; AUC: area under the curve; CT: computed tomography; DL: deep learning; ECMO: extracorporeal membrane oxygenation; ED: emergency department; ICU: intensive care unit; LoS: length of stay; LUS: lung ultrasound; ML: machine learning; PEEP: positive end-expiratory pressure; SOFA: Sequential Organ Failure Assessment.