AI-assisted quantitative lung CT evaluation following umbilical cord mesenchymal stem cell treatment in severe COVID-19 patients

Background The COVID-19 has high transmission and mortality. Previous studies support the ecacy and safety of mesenchymal stem cells (MSCs) in the treatment of lung injury. In this study, We aimed to evaluate the CT changes of lung lesions in severe COVID-19 patients treated with umbilical cord mesenchymal stem cells (UC-MSCs) by using AI-assisted quantication method. Methods 46 patients with severe COVID-19 from March 5 to April 1, 2020 were selected by single-blind, non-randomized controlled clinical study and divided into three groups: 11 cases in UC-MSCs treatment group 1 (MSC-1, with cells infusion once), 26 cases in UC-MSCs treatment group 2 (MSC-2, with cells infusion twice or three times), and 9 cases in control group with routine treatment. Repeated measure ANOVA was used to compare the effects of treatment factors on chest CT parameters of COVID-19 patients between control and experimental groups, and pairwise comparison using LSD test. The Pairwise use of immunosuppressive drugs or organ transplantation; 2. T lymphocyte abnormality (the use of allogeneic may be considered, according to the clinical opinion), HIV positive; 3. High allergic constitution or severe allergic history, especially IL-2 allergic history; 4. Pregnant and lactating women; 5. Patients with a history of serious autoimmune diseases; those who are allergic to all biological agents in the treatment, such as IL-2; 6. Patients with serious complications: Patients with chronic cardiac insuciency (NYHA cardiac function grade IV), chronic renal insuciency (CKD stage 4 or above), chronic liver insuciency (child Pugh score > 12), or malignant tumors. 7. There are other situations that the researcher thinks are not suitable for participating in this clinical study.


Introduction
In the December 2019, several cases of unexplained pneumonia emerged in Wuhan, China. By deep sequencing and virus strain isolation, it has been identi ed as a new type of acute respiratory infectious disease caused by a novel coronavirus (SARS-CoV-2) [1,2]. Of all infected patients, 4-6% developed severe pneumonia, about 5% of patients end up in ICU with acute respiratory distress syndrome, septic shock and/or multiple organ failure, and more than 75% of inpatients need oxygen [3][4][5][6]. According to the current diagnosis and treatment protocol in "Diagnosis and Treatment Protocol for Novel Coronavirus Pneumonia (Trial Version 8)", the treatment of mild to moderate patients is generally based on basic treatment, with good recovery and low mortality. However, severe patients need to be admitted to the ICU for supportive treatment such as breathing and circulation. Several studies have shown that most patients with severe infectious pneumonia will continue to develop pulmonary brosis due to poor in ammatory lesion absorption [7][8][9][10]. With the development of brotic 'scars', the lung tissue structure is gradually destroyed, especially in the alveolar area, which eventually leads to reduced gas exchange and chronic respiratory failure [11]. Although current studies have reported that dexamethasone can reduce the 28-day mortality of patients who require oxygen, and remdesivir can shorten the recovery time [12,13], no speci c treatment has been found to reduce mortality and improve the lesion absorption in convalescence.
Umbilical cord mesenchymal stem cells (UC-MSCs) is one kind of multipotent stem cells with high differentiation potential that exist in neonatal umbilical cord tissue and can be differentiated in multiple directions. Previous studies have shown that, UC-MSCs has an important role in immune regulation and damage repair. It can regulate the activation and proliferation of T lymphocytes as well as the activation and maturation of antigen-presenting cells, moreover, it can regulate the repair of airway epithelial cells after injury, alleviate the pulmonary in ammation and reduce the pulmonary brosis [14][15][16][17][18][19]. Hence, intravenous infusion of UC-MSCs is attractive therapy against severe COVID-19. Our study was aimed to use UC-MSCs to treat severe COVID-19 and evaluate the absorption of lung lesions on CT imaging by AI.

Materials And Methods
Study design.
This prospective, single-blind and non-randomized controlled clinical study was screened for COVID-19 patients in The Sixth Hospital of Wuhan (Jianghan University a liated Hospital) and Taikang Tongji (Wuhan) Hospital between March 5 and April 1, 2020. This project passed the academic review and ethical review in Shanghai East Hospital (East Hosipital a liated to Tongji University) on January 26, 2020 and January 28, 2020 respectively, and get the projects approval of the Con rmed patients at 18-75 years old were included if one meet any of the following criteria: an increased respiratory rate (≥30 times/min), an oxygen saturation≤93% when inhaled, or a PaO2 : FiO2 ≤ 300 mmHg. All patients were fully informed, agreed and signed informed consent.
Exclusion criteria: 1. Long term use of immunosuppressive drugs or organ transplantation; 2. T lymphocyte abnormality (the use of allogeneic may be considered, according to the clinical opinion), HIV positive; 3. High allergic constitution or severe allergic history, especially IL-2 allergic history; 4. Pregnant and lactating women; 5. Patients with a history of serious autoimmune diseases; those who are allergic to all biological agents in the treatment, such as IL-2; 6. Patients with serious complications: Patients with chronic cardiac insu ciency (NYHA cardiac function grade IV), chronic renal insu ciency (CKD stage 4 or above), chronic liver insu ciency (child Pugh score > 12), or malignant tumors. 7. There are other situations that the researcher thinks are not suitable for participating in this clinical study.

Randomization and masking.
Selected patients were divided into 3 groups. Principle investigators knew the group assignment for safety reasons, while participants, staff at the test site, computed tomography (CT) radiologists, and laboratory staff were unaware of the group assignment. Procedures.
The day on which the patient was treated with UC-MSCs or placebo for the rst time was de ned as D0, and 30, 60 or 90 days after that were de ned as D30, D60 or D90 respectively. Patients in control group (n=9) received intravenous infusion of placebo (100 ml normal saline containing 5% HAS) on the basis of the routine treatment that according to the treatment principle of severe and critical cases in "Diagnosis and Treatment Protocol for Novel Coronavirus Pneumonia (Trial Version 8)". Patients in MSC-1 group (n=11) received intravenous infusion of UC-MSCs (5×10 7 cells in 100 ml normal saline containing 5% HSA) for once on the basis of the routine treatment, and patients in MSC-2 group received intravenous infusion of UC-MSCs for 2 (n=10) or 3 (n=16) times, each with an interval of 5 days. The UC-MSC treatments was basically according to the clinical program in "Expert guidance on clinical research and application of stem cell treatment of COVID-19" which was joint published by the Stem Cell Biology Branch of Chinese Society for Cell Biology and Infectious Disease Branch of Chinese Medical Association. The vital signs after treatment were observed and recorded daily by the clinician. Each participant underwent a chest CT examination at D0 D30 D60 and D90 and recorded clinical scores and laboratory-related indicators.
HUC-MSCs were donations from healthy women who have undergone caesarean section, and have been approved by the ethics committee of Shanghai East Hospital, Shanghai, China. Full informed consents were obtained before delivery and donation. The manufacture and quality testing of the UC-MSC preparations used in this study were all carried out in the GMP lab of Shanghai East Hospital. Brie y, the Wharton's Jelly was aseptically separated from the umbilical cord tissue and shredded in to small pieces with sterile scissors. After shredding, the tissue was cultured in α-MEM medium (Thermo) containing 5% serum substitute (UltraGRo-Advanced). After 7-10 days, the monolayer cells crawling around the tissue block were harvested and recorded as passage 0 (P0). Cells at P5 were used for clinical treatment and comprehensive Fresh UC-MSCs were prepared once the informed consent was obtained. Quali ed UC-MSCs were suspended in 100 ml normal saline containing 5% human serum albumin (HSA) and delivered to the bedside under 2-10℃ within 8 hours. In another case, cryopreserved UC-MSC preparations which were stored in 10 mL clinical grade cryopreservation solution under -196℃ were thawed, suspended in 100 ml normal saline containing 5% human serum albumin (HSA) and delivered to the bedside under 2-10℃ within 6 hours. UC-MSCs were infused into the patients by intravenous within around 90 min.
Concerned that UC-MSC infusion may cause temporary obstruction of the pulmonary microcirculation or respiratory damage, blood pressure, heart rate, respiratory parameters and percutaneous oxygen saturation were close monitoring in real-time at the bedside before, during and 6 hours after UC-MSCs infusion.
CT image acquisition and image feature analysis.
Chest CT images were obtained from two scanners: SIEMENS SOMATOM go. Top&ALL and GE MEDICAL SYSTEMS Revolution CT with tube voltage at 130 kV or 120 kV and automatic tube current modulation at 100-400 mA. The reconstructed layer thickness is 0.8 mm or 0.6 mm, respectively. The interlayer spacing is 1.0 mm or 1.25 mm, respectively. All imaging features were reviewed and evaluated through Picture Archiving and Communication Systems (PACS, Tianjian, China) by three experienced radiologists (with 7, 8, and 28 years of chest CT diagnostic experience, respectively) with no knowledge of the group information.
CT manifestations include reticular pattern, brotic strips (irregular brotic streak shadow), halo sign, reversed halo sign, traction bronchiolectasis and structural distortion. All terms were described according to the Fleischner society glossary and the peer-reviewed literature on viral pneumonia. The distribution of lesions included peripheral, central and diffuse types. Lesions on the extra 1/3 of the lung was de ned as peripheral type, lesions on the internal 2/3 was de ned as central type, and lesions suffused through the lungs on both sides was de ned as diffuse type. Plural effusion, mediastinal and hilar lymph nodes were also evaluated.
UAI Discover-2019 nCoV (Shanghai United Imaging Intelligent Medical Co., Ltd) was used to quantitatively analyze the lung volume of patients with COVID-19. The software provided the function of segmenting lung lobes, lung segments or infection regions and quantitative analysis. After the DICOM format images of the chest CT were imported into the software, the software can quickly and automatically segment the normal areas and infection areas. Prior to this, the software has received a lot of training for COVID-19 cases in which each case was manually sketched by a senior radiologist. Based on these labeled data, the multiscale neural network model was trained to do the segmentation. In the application stage, the machine automatically extracts the lung lobe, lung segment and infected lesion area through the model after learning. The threshold range of -750 ~ -300 Hu or -300 ~ 50 Hu were de ned as GGO or Consolidation [20,21] respectively according to the segmentation results. Finally, the infection volume and percentage of the total lung, as well as the volume and percentage of the whole lung GGO, consolidation were automatically calculated by the machine ( Figure  2).

Statistical analysis.
Numerical variables were expressed as median and interquartile spacing, and intergroup comparisons were performed using one-way ANOVA and Kruskal-Wallis H tests. Categorical variables calculate frequency, and chi-square test was used for comparison. Repeated measure ANOVA was used to compare the effects of treatment factors on chest CT parameters of COVID-19 patients between control and experimental groups, and pairwise comparison using LSD test. All the statistical analysis were completed in the SPSS.25 software. The generalized linear mixed effect model analysis of ordered multiple classi cation repeated measurement data was carried out using GLIMMIX program in SAS 9.4 software. Signi cant level P <0.05.

Results
A total of 50 patients were screened with 4 patients exits, and were randomly divided into three groups. Among them, 37 cases were treated with UC-MSCs and were divided into MSC-1 group (with cells infusion once, n=11. 8 cases received fresh prepared UC-MSCs) and MSC-2 group (with cells infusion twice or three times, n=26. 8 cases received fresh prepared UC-MSCs). 9 cases with routine treatment were included in the control group ( Figure 1). The baseline characteristics of all patients, including age, sex, course of disease, clinical symptoms, comorbid diseases, clinical scores and laboratory test indicators were recorded (Table 1).

Adverse events during UC-MSCs treatment
During the entire study period, no serious adverse events related to the infusion of UC-MSCs occurred. Two patients treated with UC-MSCs developed skin itching 2 hours after infusion of UC-MSCs, and the symptoms subside the next morning.
The results indicate that UC-MSCs are safe and tolerable in the treatment of COVID-19.

Clinical and Laboratory Analyses
The clinical and laboratory results of all 46 patients are presented in Table 1. There was a signi cant difference in gender among the control group, MSC-1 group and MSC-2 group (P =0.015). 2 patients in control group had symptoms of nausea, while no cases in the other two groups (P = 0.035). The sleep score and fatigue score among three groups were signi cantly different (P = 0.006, P = 0.027). However, there was no signi cant difference among three groups in terms of age, course of disease, presence of comorbidities or the other main initial symptoms. Laboratory examination showed that, there were no signi cant differences in the white-cell count, lymohocyte count, monocyte count, neutrophil count, Creaction protein, D-Dimer, PaO 2 or PaCO 2 among three groups.

Chest CT Analysis on Admission
The chest CT ndings of patients were presented in Table 2. There were no signi cant differences in the lesion distribution and processing data based on arti cial intelligence among three groups.

Effects of stem cells on CT parameters during follow-up
Repeated measures ANOVA was used to compare the imaging features of control group and MSC groups in 30-90 days follow-up. The results were shown in Table 3. Within 30-90 days of follow-up, the differences between the percentage of GGO in total lung or the percentage of total lung infection volume on day 0 and that on the three follow-up time points showed a decreasing trend with time. However, the differences between the percentage of consolidation on day 0 and that on the three follow-up time points shows no signi cant downward trend and there was no signi cant difference between the control group and the two MSC treated groups. (Figure 3).
Repeated measures ANOVA was used to compare the imaging features of control group and MSC groups in 60-90 days follow-up. The results were shown in Table 4. The differences between the percentage of GGO in total lung or the percentage of total lung infection volume on day 0 and that on day 60 as well as on day 90 showed a decreasing trend with time. The results of pairwise comparison showed that the differences between the percentage of GGO in total lung or the percentage of total lung infection volume on day 0 and that on day 60 as well as on day 90 in MSC-1 group were smaller than those of control group and MSC-2 group, but there was no statistical signi cance between control group and MSC-2 group.

Effects of stem cells on CT structural characteristics during follow-up
A generalized linear model was used to analyze the dynamic changes of CT structural characteristics between the control group and MSC groups within 90 days follow-up. The results are shown in Table 5. There were no statistically signi cant differences in the terms of traction bronchiolectasis, architectural distortion and reticular pattern among three groups. It can be concluded that UC-MSC treatment have no effect on the dynamic changes of these three CT structures. The three structures all changed over time (P <0.001).

Discussion
The results of our non-randomized controlled single-blind clinical trial showed that, severe COVID-19 patients who received UC-MSC infusion once had a statistical signi cant difference in the absorption of lung CT lesions at 60 days and 90 days compared with patients in control group. However, UC-MSC infusion 2 or 3 times not only did not improve the e cacy, but appeared to worsen the prophetic e cacy. This might be mainly due to the usage of freshly thawed cryopreserved cells which will be discussed in details below. By the way, the AI-assisted evaluation method of lung CT lesions is more e cient and comparable.
For severe COVID-19 patients, after a comprehensive treatment in the acute phase, pulmonary brosis will also cause longterm functional damage in the convalescence. Currently, there was no speci c treatment in the world. Many researchers have studied the pathogenesis of severe cases and believe that in ammatory storm may be one of the key causes of serious lesions in patients [22][23][24], so immunomodulatory therapy was considered and MSCs was chosen to treat severe COVID-19 due to its excellent immunomodulatory ability. Zhang et al. performed MSC treatment for one COVID-19 patient.
The lung function improved 2 days after MSC transplantation and was discharged 7 days later, emphasizing the potential of stem cells in the treatment of COVID-19 [25]. Zhao Chunhua's group used UC-MSCs to treat 7 patients with COVID-19. Their results con rmed that UC-MSCs were safe for COVID-19 patients and can promote the absorption of local lung lesions and reduce multiple in ammation indicators [26]. Recently, Wang Fusheng's group published a phase I clinical study of stem cell treatment of COVID-19. A total of 18 patients were enrolled, 9 of which were treated with UC-MSCs. Their results showed that the three-times infusion of UC-MSCs was safe and well tolerated, except for two patients experienced transient facial ushing and fever and one patient experienced transient hypoxia 12 hours post UC-MSCs infusion. No serious adverse events related to UC-MSCs infusion were observed [27]. Compared with the above three studies, our study expanded the number of patients enrolled. A total of 37 severe COVID-19 patients were treated with UC-MSCs, and no serious adverse events were observed. The baseline imaging signs, total lung infection percentage, lung GGO and consolidation volume of the control group and UC-MSC treatment groups (MSC-1 and MSC-2) were similar with no statistical difference. Compared with the control group, the absorption of CT lung lesions at 60 days and 90 days was more obvious in MSC-1 group, and the difference was statistically signi cant. However, there were no signi cant difference in the absorption of lung lesions at 60 days and 90 days between the MSC-2 group and the control group. In addition, the MSC-1 group had more obvious lung lesion absorption than the MSC-2 group. The possible reasons analyzed were as follows: Firstly, in our study, 8 of the 11 patients in the MSC-1 group were treated with fresh prepared UC-MSCs, however, 8 of the 26 patients in the MSC-2 group received fresh prepared UC-MSCs only for the rst infusion and others all received cryopreserved MSCs for twice or three times, which suggested that cryopreserved MSCs which were freshly thawed may have certain defects in terms of e cacy. In 2009, a multi-center phase III clinical trial used industrial bone marrow mesenchymal stem cell (BM-MSC) products to treat GvHD, but did not reach its main clinical endpoint. Compared with placebo control, there was no signi cant improvement of the patients in the BM-MSC treatment group and the 180-day mortality rate of the two groups was equivalent [28]. When analyzing the causes [29], the investigator pointed out that the infusion of cryopreserved MSC immediately after thawing maybe an important reason to cause clinical ineffectiveness.
Studies found that cryopreserved human MSCs started a heat shock program within 4 hours post-thaw, and signi cantly reduced the immunosuppressive properties that depend on indoleamine2,3-dioxygenase (IDO) [30], and its metabolic activity was also signi cant decrease [31]. Compared with fresh prepared MSCs, the cytoskeleton protein F-actin was reduced by 60% in cryopreserved MSCs, and its ability to bind to human endothelial cells was reduced by 80%, and its survival time after infused into mice was signi cantly shortened [32]. However, this decline in immunosuppressive ability post-thaw can be recovered by culturing in vitro for 24 hours [30]. But it was still di cult to predict whether this function recovery can be carried out in vivo, because even "live" MSCs may be eliminated due to abnormal metabolic function [33]. Therefore, we speculate that patients in the MSC-2 group did not show e cacy, which may be caused by the infusion of cryopreserved MSCs with decreased activity. This also suggests that in future clinical applications, cryopreserved stem cells need to be cultured for 24-48 hours pre-transplantation to restore their optimal immune regulation and biological activity.
Furthermore, as Killer et al. reported [31], the immunosuppressive ability of the same batch of MSCs to T cells from different donors was quite different. Therefore, the individual differences in severe COVID-19 patients may also be the cause for the difference in e cacy between the MSC-1 and MSC-2 groups. The in uence of such factors can be reduced by expanding the sample size.
AI-assisted evaluation of pneumonia lesions has been reported in a large number of literatures[34-37]. The currently accepted conclusion is that the accuracy of AI used to evaluate COVID-19 is no worse than manual evaluation, and the in uence of arti cial subjective factors is excluded [38]. These previous studies, especially our research aimed to make a comparison of the absorption of lung lesions before and after treatment, indicate that AI-assisted evaluation was more comparable. We believe that the results of AI evaluation were more reliable than manual evaluation, so we did not compare the differences between AI and manual evaluation.
The pathological changes of pulmonary brosis and the absorption of brotic lesions require a long-term process.
Whether UC-MSCs treatment has a positive effect on slowing down the development of pulmonary brosis in patients still needs long-term observation. Therefore, we are continuing to follow up these patients.
Limitations of this study: 1, the research design did not strictly follow the 1:1 allocation of the UC-MSC group and the control group, resulting in fewer cases in the MSC-1 group and the control group. 2, only the imaging results of the followup patients were analyzed, and the clinical laboratory and other indicators were not integrated. 3, there were only 3 cases of once infusion with cryopreserved UC-MSCs, therefore, it was impossible to make further statistical analysis on the difference between freshly prepared UC-MSCs and cryopreserved UC-MSCs in the absorption of lung lesions. Our future research will expand the sample size, implement a strict randomized controlled design, and combine multiple indicators to evaluate the safety and e cacy of UC-MSCs in the treatment of COVID-19.

Conclusion
For severe COVID-19 patients, the infusion of UC-MSCs can improve the absorption of lung lesions with no serious adverse events. AI is more suitable for the comparative evaluation of CT lung lesions before and after treatment.         Chest CT scanned images of a 67-year-old men in the control group at D0 and D90. A&B, the red area was the infection area automatically identi ed by AI software. C&D are original images of A and B respectively indicated that the infection areas identi ed by AI software were quite accurate. E, the change trend of the percentage of lung infection in the two examinations at D0 and D90.