Lung CT scan has been used extensively for more than 20 years to improve understanding of the pathophysiology of ARDS. In particular, the quantitative analysis of the ARDS lung CT scan has allowed the quantification of the amount of not aerated tissue, poorly aerated tissue, well-aerated tissue, and over inflated tissue, advancing the concept of the baby lung and of the lung as a “sponge model” ^[1][26]. This quantitative approach has shown the redistribution of the densities in prone position and the change in the not aerated tissue fraction at two airway pressures considered the gold standard for assessing recruitment in ARDS. In this context, CT scan has represented a useful tool to decide the better mechanical ventilation strategy ^[2][27]. Similarly, in the setting of chest trauma, the quantification of parenchyma damage, at hospital admission by CT scan, can help predict the evolution from the initial traumatic injuries to focal or diffuse alveolar hemorrhage followed by pulmonary edema and interstitial alterations typical of ARDS ^[3][28].

Then, recently during the COVID-19 pandemic, CT scan has played a relevant role as a screening tool due to its greater sensitivity for detecting early pneumonic changes. In fact, although COVID-19 is typically confirmed by viral nucleic acid detection, lung CT scan has been widely used to differentiate COVID-19 from other viral pneumoniae and to predict the severity of pneumonia even in the early stage ^[4][29].

The application of AI on CT scan lung images has been recently implemented in patients with COVID-19 disease to predict the evolution in ARDS, to stage and quantify the disease, and predict the outcome.

Similarly, in non-COVID-19 ARDS, AI has the potential to give a profound contribution, considering its ability to automatically and efficiently analyze and segment acutely injured lungs, to provide automated quantitative analysis, and to predict the development of ARDS, the alveolar recruitment, and the relationship between the quantitative analysis of lung tissue and specific outcomes.

The diagnosis and prediction of ARDS have been supported by various systems, tools, and techniques, both before and after the AI revolution. Imaging techniques such as chest radiography, CT, and LUS have played a critical role in the diagnosis and management of ARDS ^[5][6][30,31]. These lung imaging modalities have been essential in assessing lung aeration, predicting oxygenation response, and facilitating early diagnosis to prevent the progression of lung injury ^[7][32]. In addition, biomarkers have been explored for their potential in diagnosing ARDS and predicting its prognosis ^[8][9][33,34]. Recent advances in AI have significantly changed the landscape of traditional imaging techniques, allowing for more accurate and rapid analysis of medical imaging data for ARDS diagnosis and severity prediction ^[10][35].

AI-based diagnostic models combining clinical data and CT scans have been developed, providing accurate and explainable ARDS diagnostic models for real-life scenarios ^[11][12][36,37].

Recently, AI has facilitated the development of models to predict subsequent ARDS development using features identified at initial presentation with COVID-19, addressing the need for clinical decision support tools during the early stages of the COVID-19 pandemic ^[13][14][38,39].

In addition to COVID-19 pneumonia, blunt chest trauma is also currently associated with parenchymal lung injury to various extents, which may increase the risk of developing ARDS. Typically, the presence of an injury severity score (ISS) greater than 25 significantly increases the risk of developing ARDS ^[15][18]. Moreover, it has previously been shown that trauma patients with pulmonary contusions involving at least 20% of the total lung volume have a significantly higher risk of developing ARDS ^[16][40]. Thus, the ability to assess early information on lung CT that may be associated with the risk of developing ARDS could allow for timely supportive therapy. Röhrich et al. developed a ML method for the early prediction of ARDS based on CT in trauma patients at hospital admission. One hundred and twenty-three patients were enrolled. The model consisted of a fully automated ML and radiomics-based approach that showed a higher accuracy compared to an established score (ISS and abbreviated injury score of the thorax) to identify ARDS in trauma patients ^[15][18]. In this line, a rapid automated lung CT volumetry assessment of pulmonary contusions in trauma patients showed a good accuracy in assessing the risk of ARDS, length of intensive care stay, and time on mechanical ventilation ^[17][19].

ARDS is characterized by widespread inflammation in the lung, leading to increased permeability of the alveolar–capillary barrier, impairment of pulmonary mechanical properties, and impaired gas exchange. There is often a phenomenon known as alveolar collapse or atelectasis, where some of the alveoli collapse and are not involved in gas exchange ^[18][19][41,42].

Mechanical ventilation itself can increase or cause lung damage known as ventilator-induced lung injury (VILI). Therefore, the therapeutic goal of mechanical ventilation in ARDS patients is not only to maintain “normal gas exchange” but also to protect the lung from VILI ^[20][21][22][43,44,45].

Lung protective strategies include low tidal volumes and adequate levels of positive end-expiratory pressure (PEEP) levels to keep the alveoli open and prevent them from collapsing. However, too high or too low levels of PEEP can lead to damage due to overdistension or cyclic collapse. Estimation of the percentage of lung that can be recruited or re-opened by applying transient increases in airway pressure has been demonstrated to be associated with the response to PEEP and prone position ^[2][23][27,46]. Thus, the quantification of alveolar recruitment can help the clinician optimize the protective ventilation strategy to avoid VILI.

The introduction of the lung CT quantitative analysis has allowed the assessment and quantification of the aerated and not aerated lung regions and the possible changes due to the mechanical ventilation and body position ^[24][6]. In particular, the application of the lung CT quantitative analysis performed at two different levels of airway pressure is considered the gold standard for the assessment of alveolar recruitment, because it can calculate the difference of not aerated tissue ^[25][26][27][47,48,49].

However, since the mid-1980’s, the application of quantitative analysis was rarely used in clinical practice because it requires the manual segmentation of the lung by the physicians ^[26][48]. The assessment of lung recruitment can take up to 6–8 h with a certain degree of error ^[28][15]. To improve the ability to assess lung recruitment, a visual anatomical evaluation of recruitment has been proposed ^[26][48].

Moving from the successful application of DL to the segmentation process of CT lung images in ARDS ^[29][30][50,51], using two CNNs architectures, the Seg-Net and the U-Net, Herrmann et al. decided to implement the U-net to develop a DL algorithm to automatically segment injured lungs affected by ARDS and to calculate lung recruitment by performing two CT scans at 5 and 45 cmH₂O of airway pressure ^[31][11].

Training was performed on 15 healthy subjects (1302 slices), 100 ARDS patients (12,279 slices), and 20 COVID-19 patients (1817 slices): 80% of the patients were used for training and 20% for testing. The authors found that automatic lung segmentation performed by a properly trained neural network was reliable and closely matched the results obtained by manual segmentation. In fact, the total lung volume measured by AI and manual segmentation had a R² of 0.99 and a bias of −9.8 mL (CI +56.0/−75.7 mL). Although the model was not perfect, especially in the most damaged lung areas, which are difficult to identify even for a trained radiologist, but which did not exceed 10% of the lung parenchyma, the AI segmentation showed the same degree of inaccuracy as the manual segmentation. In fact, for recruitability measured using manual and AI segmentation, change in not aerated tissue fraction had a bias of +0.3% (CI +6.2/−5.5%) while −0.5% (CI +2.3/−3.3%) was expressed for change in well-aerated tissue fraction.

Subsequently, Penarrubia et al. in a single center study assessed both intra- and interobserver smallest real difference exceeding measurement error of recruitment using both human and ML lung segmentation on CT scan ^[28][15]. Low-dose CT scans were acquired at 5 and 15 cm H₂O of PEEP in 11 sedated and paralyzed ARDS patients and recruitment was computed as the change in weight of the not aerated lung regions. The intra-observer small real difference of recruitment was 3.5% of lung weight, while the human–human interobserver smallest real difference of recruitment was slightly higher amounting to 5.7% of lung weight, as also was the human–machine smallest real difference. Human–machine and human–human interobserver measurement errors were similar, suggesting that ML segmentation algorithms are a valid alternative to humans for quantifying alveolar recruitment on CT ^[28][15].

Furthermore, to overcome the difficulty in performing two CT scans at two different airway pressures, Pennati et al. developed a ML algorithm to predict lung recruitment in ARDS patients, starting from a single CT scan obtained at 5 cmH₂O upon admission to the intensive care unit (ICU) ^[32][2].

The authors demonstrated that in 221 retrospectively analyzed ARDS patients, the use of four ML algorithms (logistic regression, support vector machine, random forest, XGboost) based on a lung CT scan at 5 cmH₂O were able to classify lung recruiter patients with similar area under the curve (AUC) compared to a ML model based on the combination of lung mechanics, gas exchange, and CT data ^[32][2].

The application of this ML algorithm with an automatic lung segmentation and quantitative analysis could reduce the workload and ionizing radiation exposure of the traditional method of assessing lung recruitability.

Concerning the outcome, hospital mortality has decreased over the decades, but has remained unchanged in recent years, despite advances in supportive care ^[33][52].

A small retrospective study of 42 patients with ARDS evaluated the relationship between the volume of well-aerated lung regions, calculated automatically by software, and outcome ^[34][9]. Total lung volumes and well-aerated lung regions were significantly higher in survivors. Estimates of the total volumetry and the regions of interest were obtained within three minutes with a very good reproducibility ^[34][9].

Several data have shown that lung volume and the amount of not aerated lung areas in COVID-19 are associated with respiratory severity and outcome ^[35][53]. Typical lung CT findings in COVID-19 patients include bilateral pulmonary ground-glass opacities and opacities with rounded edges usually localized in the peripheral lung regions ^[24][6].

Using a DL method to calculate the description of the CT, two clusters typically associated with COVID-19 and two clusters associated with bacterial pneumonia were found ^[36][12]. The clusters containing diffuse ground-glass opacities in the central and peripheral lung showed up to 91% accuracy in correctly classifying COVID-19 and pneumonia.

Liu et al. investigated the ability of quantitative lung CT analysis compared to traditional clinical biomarkers to predict progression to severe disease in the early stage of COVID-19 patients ^[37][14]. A group of 134 patients with COVID-19 who underwent lung CT scan and laboratory tests on day 0 and 4 were enrolled. All patients were followed up for 28 days until the first occurrence of severe disease or otherwise. Three AI-derived CT features were calculated according to Hounsfield units (−700/−500; −500/−200; and −200/−60 HU). The CT features at day 0 and day 4 and their changes from day 0 to day 4 showed the best discriminative ability to predict patient progression to severe disease. In this line, a retrospective study of COVID-19 patients used DL segmentation to assess lung volume and density composition ^[38][20]. The number of lung regions with a density between −549 and −450 of Hounsfield units was associated with an increased risk of ARDS. Although the results were not published, Lopes et al. proposed a multicenter retrospective longitudinal study to correlate the possible findings on lung CT in patients with COVID-19 infection and the course of the disease ^[39][16].

In COVID-19 patients, the use of the quantitative lung CT analysis at hospital admission, which calculates the volume of the affected lung as the sum of the poorly aerated and not aerated lung regions, predicted the need for oxygen support and intubation with good accuracy ^[40][13].

Regarding hospital mortality in COVID-19 ARDS based on AI quantification of lung involvement at hospital admission, the AI did not predict the outcome ^[41][17]. In contrast, the Sequential Organ Failure Assessment (SOFA) score resulted in an AUC for hospital mortality of 0.74 (95% CI 0.63–0.85), suggesting that other clinical parameters reflect the overall disease severity ^[42][43][54,55].

In severe COVID-19 ARDS patients with hypoxemia refractory to the conventional ventilation, veno-venous extracorporeal membrane oxygenation (ECMO) may be used to improve the outcome. However, the potential improvement in outcome is higher when ECMO support is applied in the early phase. Therefore, a possible early stratification should be considered. The use of an AI-based quantification of lung involvement was able to predict the need for ECMO with an acceptable AUC [0.83 (95% CI 0.73–0.94)] ^[44][10]. In addition, combining the SOFA score with CT lung involvement at ICU admission improved the AUC to 0.91 (95% CI 0.84–0.97) ^[44][10].

In summary, while the extent of lung involvement on imaging is an important consideration in assessing the severity of ARDS, there is not a specific “critical amount” that universally predicts outcome or guides the need for repeat imaging or ECMO use ^[44][10]. The decision to use ECMO is multifactorial and is based on clinical judgment, including considerations of the patient’s overall health, the underlying cause of ARDS, and the potential for recovery. Similarly, the timing and need for repeat imaging are individualized based on the clinical course of the patient and the judgment of the healthcare provider. In fact, the early clinical course of the disease may be more predictive of the outcome than the assessment at time of admission to the ICU ^[45][56].