Key words
COVID-19 - computed X-Ray tomography - pneumonia - artificial intelligence - lung
volume measurements - laboratory tests
Introduction
The novel coronavirus SARS-CoV-2 has spread from Wuhan city, China to most other countries
worldwide. More than one year after the beginning of the pandemic, many European countries
are facing a “fourth wave”. Although no longer recommended as a first-line test, CT
of the chest still plays an important role in the diagnostic workup and follow-up
of Coronavirus Disease 2019 (COVID-19) [1]. The typical CT findings in COVID-19 pneumonia are multiple ground-glass opacities
and consolidation predominantly located in the periphery of the mid and lower lung
zones [2]
[3]
[4]. It has been shown that the burden of pulmonary opacifications correlates with prognosis
[5]
[6]
[7]. While a rough assessment of pulmonary involvement can be done by subjective eyeballing,
human assessment is not as reliable in exact grading [8]. Computer-assisted quantitative analysis of CT images might facilitate and accelerate
the precise assessment of pulmonary changes related to COVID-19 pneumonia. Also, quantifiable
results might enable the calculation of accurate scores, for example, regarding individual
prognosis. Furthermore, reader variability could be diminished.
To date, many studies have investigated the application of artificial intelligence
(AI) in COVID-19, but only a few have examined AI-based analyses of CT images to predict
prognosis [8]
[9]
[10]
[11]
[12]
[13]
[14]. These studies used automatic or semiautomatic procedures, different radiological
parameters, and different segmentation techniques. As expected, it was confirmed that
the extent of lung opacities was a prognostic marker for an unfavorable clinical outcome.
Two studies only distinguished between different densities of opacifications [9]
[14]. However, this differentiation might be important because several studies demonstrated
that ground glass opacities are the predominant finding in patients with a favorable
outcome, whereas consolidation predominates in cases with an adverse outcome [15]
[16]. Other groups investigated and determined the prognostic relevance of a number of
different laboratory values [17]
[18]
[19]
[20]. Zheng et al. [21] evaluated a combined model of clinical and subjective CT features for the prediction
of an adverse outcome in patients with COVID-19, while Schalekamp et al. [22] developed a predictive risk model for critical illness in patients with COVID-19
based on laboratory findings and visual, semi-quantitative analysis of chest radiographs.
The potential benefit of a combined model of quantitative analysis of CT images and
laboratory values has not yet been evaluated. However, combining a marker of actual
lung involvement with a marker of systemic disease seems reasonable for determining
severity of disease.
The aim of our study was to evaluate if the AI-based analysis of chest CT images stratified
according to density and extent of opacifications correlates with the clinical outcome
of patients suffering from COVID-19 pneumonia. Furthermore, we evaluated if a combined
model including CT findings and laboratory data enhances prognostic power.
Materials and methods
This study was approved by the institutional ethics committee. Written informed consent
was waived due to the retrospective nature of this study.
Patient recruitment
Cohort for model development
120 consecutive patients from one academic tertiary care hospital and one general
tertiary care hospital in Bavaria, Germany with COVID-19 infection confirmed by Polymerase
Chain Reaction (PCR) were included in a first step for model development. Chest CT
scans were acquired between March 07 and April 27, 2020. The patients had been admitted
for CT scan either on first admission or after transfer from another hospital. The
indication for the CT scan was based on clinical assessment. In case of more than
one CT per patient, only the first CT scan was considered for analysis. Exclusion
criteria were a negative chest CT for COVID-19 pneumonia (2 patients), admission to
intensive care unit (ICU) > 3 days before CT scan (9 patients), pneumothorax (1 patient),
incomplete CT scan data (3 patients) and incorrect automatic lung segmentation (16
patients). Thus, 89 patients were finally included into the analysis.
Validation cohort
The validation cohort consisted of 32 patients from the same academic tertiary care
hospital with PCR-confirmed COVID-19 infection with the same exclusion criteria as
the first cohort. Chest CT scans were acquired between May 05, 2020 and November 13,
2020. Again, only the first CT scan from each patient was considered for analysis.
One patient was excluded due to incorrect automatic lung segmentation, 3 patients
were excluded due to admission to ICU > 3 days before CT scan. Thus, 28 patients were
ultimately included in the validation cohort.
The flowchart of patient inclusion is shown in [Fig. 1].
Fig. 1 Patient selection process.
Abb. 1 Patientenselektion.
Epidemiology and laboratory data
Patient characteristics (age, sex), date of symptom onset, and laboratory data were
extracted from electronic patient records. With respect to laboratory data, C-reactive
protein (CRP), white blood cell count (WBC), relative lymphocyte count (RLC), relative
eosinophil count (REC), troponin, NT-proBNP, fibrinogen, interleukin-6 (IL-6), D-dimer,
lactate dehydrogenase (LDH), creatine kinase (CK), CK-MB, and lactate were included
in the analysis. Laboratory values were recorded as close as possible to the date
of the CT scan, but not more than 3 days before or after.
Definition of outcome
For analysis of prognosis, two compound outcomes were defined: positive outcome was
defined as either discharge or regular ward care; negative outcome was defined as
need for mechanical ventilation, treatment in the ICU, extracorporeal membrane oxygenation
(ECMO), or death.
Image acquisition
CT scans were performed using multi-detector spiral CT scanners (SOMATOM Definition
Flash/SOMATOM Sensation 16, Siemens Healthcare GmbH) in supine position during end-inspiration.
Intravenous contrast material was administered at the discretion of the radiologist
considering the individual study indication (e. g., if pulmonary embolism or superinfection
was suspected). Automatic tube voltage selection was applied with a reference tube
voltage of 120 kV. The tube current was regulated by an automatic tube current modulation
technique with the reference mAs being 40–110. The tube current was 100 kV and 140 kV
when the dual-energy technique was applied. Collimation width was 0.625 mm–0.75 mm.
Axial planes were reconstructed with a slice thickness of 0.75–1.5 mm (81 CTs) and
2–5 mm (8 CTs) in the lung kernel and with a slice thickness of 0.75–1 mm (82 CTs)
and 2–5 mm (7 scans) in the soft-tissue kernel using either filtered back-projection
or iterative reconstructions. Dual-energy acquisitions were reconstructed as a weighted
average with 60 % from the 140 kVp and 40 % from the 100 kVp data. Additional sagittal
and coronal MPRs were reconstructed with a slice thickness of 1–3 mm using a lung
and a soft-tissue kernel. The pictures were sent to a picture archiving and communication
system (PACS, Syngo Imaging, Siemens Healthcare GmbH).
Image evaluation
Axial reconstructions of the CT scans in the soft-tissue kernel were processed in
syngo.via (Siemens Healthcare GmbH) using the CT Pneumonia Analysis prototype (CTPA).
This software performs automatic lung segmentation and differentiates between normal
lung tissue and opacified lung tissue by a deep learning-based algorithm. This software
has been validated before including CT scans with a slice thickness of > 3.0 mm [23]. Opacified lung tissue is subdivided into two categories, namely density below and
above a cutoff-value of –200 HU. This threshold is preset by the software and represents
the approximate transition of ground glass to consolidation. The automatic segmentation
of the lung parenchyma and opacities was reviewed for correctness by a senior radiologist
who was blinded to clinical data and laboratory data. The software calculated the
following values: lung volume (LV, ml), volume of opacity (VO, ml), percentage of
opacity (= VO/LV, PO, %), volume of high (> -200HU) opacity (VHO, ml), percentage
of high opacity (= VHO/LV, PHO, %), and mean HU of the whole lung (MHUL, HU).
Statistical considerations
Sample size
For both cohorts, the maximal available sample size within the predefined periods
was included. With a sample size of 89 patients with 38 patients experiencing a negative
outcome, a prediction model with 3 predictors had a sufficient number of events per
variable (EPV)[24] to get a stable and not overfitted regression model. The validation cohort with
28 patients (10 negative outcomes) was large enough to give first evidence about the
general performance of the developed score.
Statistical methods
Continuous variables are presented as means ± standard deviation for normal distributed
or as median (q1-q3) for non-normal distributed variables. Categorical variables are
presented as absolute numbers and percentages. Univariable logistic regression analyses
were performed to identify predictors for an unfavorable outcome. A multivariable
logistic regression model was built according to the following criteria: 1) only predictors
with a p-value < 0.1 in the univariable model and 2) with no more than 10 % missing
values were added. 3) If two predictors were highly correlated with r > 0.8 (multicollinearity),
only one was included in the model. 4) A backward selection was performed and only
predictors with p < 0.05 remained in the final model. Odd’s ratios (OR) and corresponding
95 % confidence intervals are presented as effect estimates for all logistic regression
models. ORs are presented as per 1 unit change for each predictor, if not otherwise
specified. Finally, a prediction score was established by dichotomizing each predictor
of the multivariable model. The optimal cut-off for each predictor was chosen according
to the Youden-Index in the univariable case.
Model validation
The performance (calibration and discrimination) of the final score was evaluated
by a calibration plot and the concordance statistic (c-index). An internal validation
of the score to obtain stable optimism-corrected estimates was performed using bootstrap
validation with n = 1000 replications. The model was further validated on the validation
cohort by calculating the score for each patient and assessing the predictive accuracy
for a negative outcome. A p-value < 0.05 was considered statistically significant
for all analyses. All analyses were performed using R, version 4.1.1 (The R Foundation
for Statistical Computing).
Results
Patient data, outcome
Cohort for model development
A total of 89 CT examinations were evaluated. The average age of the patients was
60.3 years (SD 14.4). 35 (39.3 %) of the 89 patients were female. 68 CT scans were
performed without contrast agent, 21 were contrast-enhanced for clinical reasons.
The mean time period from the onset of symptoms to the CT scan was 9.6 days (SD 5.7).
51 patients (57.3 %) had a positive outcome, 38 patients (42.7 %) a negative outcome.
The average timespan from CT scan to negative outcome was 1.6 days (SD 2.5; –1–8 days).
Validation cohort
28 CT scans were evaluated. The average age of the patients was 53.8 years (SD 13.4).
9 (32.1 %) patients were female. 7 CT scans (25 %) were contrast-enhanced. The mean
time from the onset of symptoms to the CT scan was 7 days (SD 3.3). 18 patients (64.3 %)
had a positive outcome, 10 patients (35.7 %) a negative outcome. The average timespan
between CT scan and negative outcome was 0 days (SD 1.6; -3–3 days).
Patient data is summarized in [Table 1].
Table 1
Patient demographics.
Tab. 1 Demografische Daten.
|
All patients
|
Positive outcome
|
Negative outcome
|
p
|
|
Cohort for model development
|
|
Number of patients
|
89 (100)
|
51 (57.3)
|
38 (42.7)
|
|
|
Female
|
35 (39.3)
|
21 (41.1)
|
14 (36.8)
|
0.68
|
|
Age
|
60.3 ± 14.4
|
60.8 ± 15.2
|
59.7 ± 13.5
|
0.72
|
|
Time between symptom onset and CT (days)
|
9.6 ± 5.7
|
10.0 ± 6.0
|
8.8 ± 5.4
|
0.43
|
|
Time between CT and negative outcome (days)
|
|
|
1.6 ± 2.5
|
|
|
Validation cohort
|
|
Number of patients
|
28 (100)
|
18 (64.3)
|
10 (35.7)
|
|
|
Female
|
9 (32.1)
|
6 (33.3)
|
3 (30)
|
0.91
|
|
Age
|
53.8 ± 13.4
|
53.1 ± 15.1
|
55.2 ± 10.4
|
0.72
|
|
Time between symptom onset and CT (days)
|
7 ± 3.3
|
6 ± 3.2
|
8.7 ± 3.1
|
0.06
|
|
Time between CT and negative outcome (days)
|
|
|
0 ± 1.6
|
|
Data are presented as n (%) or mean ± standard deviation; positive outcome: discharge
or regular ward care; negative outcome: need for mechanical ventilation, treatment
in intensive care unit, ECMO, or death.
Univariable logistic regression analyses
Quantitative CT analysis
The results of quantitative CT analysis are summarized in [Table 2]. All of the 6 quantitative CT values were significantly associated with a negative
outcome.
Table 2
Predictive value of quantitative CT analysis on patient outcome.
Tab. 2 Prädiktive Aussagekraft der quantitativen CT-Analyse bezüglich des Outcomes.
|
Variable
|
All patients
|
Positive outcome
|
Negative outcome
|
OR (95 %-CI)
|
p
|
|
LV (ml)
|
4083 (3049–4984)
|
4486 (3594–5187)
|
3461 (2739–4408)
|
0.94 per 100 ml (0.90–0.98)
|
0.001
|
|
VO (ml)
|
604 (293–1496)
|
427 (173–923)
|
1423 (671–1936)
|
1.16 per 100 ml (1.08–1.25)
|
< 0.001
|
|
PO (%)
|
14.4 (6.8–47)
|
10.0 (4.0–16.5)
|
49.4 (18.5–61.4)
|
1.07 per 1 % (1.04–1.10)
|
< 0.001
|
|
VHO (ml)
|
79 (32–237)
|
53 (13–110)
|
242 (75–537)
|
1.92 per 100 ml (1.34–2.74)
|
< 0.001
|
|
PHO (%)
|
2.2 (0.7–7.6)
|
1.2 (0.3–3.0)
|
7.8 (2.1–16.3)
|
1.23 per 1 % (1.02–1.39)
|
< 0.001
|
|
MHUL
|
–718 (–784–609)
|
–764 (–794–684)
|
-599 (–718–513)
|
1.13 per 10 units (1.07–1.20)
|
< 0.001
|
LV: lung volume; VO: volume of opacity; PO: percentage of opacity; VHO: volume of
high opacity; PHO: percentage of high opacity; MHUL: mean HU of the whole lung; MHUO:
mean HU of opacity. Data are presented as median (q1-q3); positive outcome: discharge
or regular ward care; negative outcome: need for mechanical ventilation, treatment
in intensive care unit, ECMO, or death; OR: odds ratio; CI: confidence interval.
Laboratory values
A significant association with a negative outcome was found for CRP, RLC, troponin,
and LDH. The laboratory values and their predictive value regarding patient outcome
are summarized in [Table 3].
Table 3
Predictive value of laboratory values on patient outcome.
Tab. 3 Prädiktive Vorhersagekraft der Laborwerte auf das Outcome.
|
Laboratory value
|
Number of patients with value available (%)
|
All patients
|
Positive outcome
|
Negative outcome
|
OR (95 %-CI)
|
p
|
|
CRP (mg/l)
|
86 (97)
|
61.7 (31.7–145)
|
40.3 (18–61.3)
|
130 (85–195)
|
1.01 (95 %-CI:1.01,1.02)
|
< 0.001
|
|
WBC (cells/nl)
|
85 (96)
|
6.6 (4.7–8.2)
|
5.9 (4.4–7.2)
|
7.2 (5.4–10.2)
|
1.02 (95 %-CI:0.99,1.07)
|
0.4
|
|
RLC (%)
|
83 (93)
|
13.6 (9.9–20.6)
|
18.1 (12.1–25)
|
11.7 (9–14.4)
|
0.95 (95 %-CI:0.9,0.99)
|
0.05
|
|
REC (%)
|
80 (90)
|
0.1 (0–0.8)
|
0.1 (0–0.8)
|
0 (0–0.6)
|
0.98 (95 %-CI:0.67,1.39)
|
0.91
|
|
Troponin (ng/l)
|
47 (53)
|
12 (7.3–34.5)
|
8.4 (6–14)
|
30 (8–69.3)
|
1.06 (95 %-CI:1.02,1.12)
|
0.03
|
|
NT-proBNP (pmol/l)
|
35 (39)
|
387 (138.5–1112)
|
224 (115–469)
|
455 (186–2576)
|
1 (95 %-CI:1,1)
|
0.19
|
|
Fibrinogen (mg/dl)
|
27 (30)
|
579.3 (512.5–645.5)
|
601.2 (588–614.3)
|
579.3 (504–650)
|
1 (95 %-CI:0.99,1.01)
|
0.85
|
|
IL-6 (pg/ml)
|
36 (40)
|
90.4 (34.4–186.8)
|
35.1 (30.7–62.9)
|
129.8 (74.7–233.4)
|
1.01 (95 %-CI:1,1.03)
|
0.06
|
|
D-dimer (mg/l)
|
68 (76)
|
284 (12.9–970.8)
|
263 (154.5–578.3)
|
393.5 (10.5–1083)
|
1 (95 %-CI:1,1)
|
0.11
|
|
LDH (U/l)
|
82 (92)
|
336.5 (262–473.5)
|
289 (243.5–356)
|
485 (360.5–611.5)
|
1.01 (95 %-CI:1.01,1.02)
|
< 0.001
|
|
CK (U/l)
|
53 (60)
|
145 (73–309)
|
89 (57.5–243.3)
|
186 (132–436)
|
1 (95 %-CI:1,1.01)
|
0.06
|
|
CK-MB (ng/ml)
|
23 (26)
|
9.4 (1.5–18.2)
|
2.3 (1.4–7.9)
|
11.4 (3.4–20.1)
|
1.07 (95 %-CI:0.97,1.21)
|
0.2
|
|
Lactate (mmol/l)
|
73 (82)
|
0.16 (0.09–0.89)
|
0.12 (0.09–0.83)
|
0.18 (0.09–1)
|
1.01 (95 %-CI:0.94,1.09)
|
0.8
|
Data are presented as median (q1-q3); positive outcome: discharge or regular ward
care; negative outcome: need for mechanical ventilation, treatment in intensive care
unit, ECMO, or death; OR: odds ratio; CI: confidence interval.
Multivariable logistic regression analysis
The final multivariable logistic regression model after variable selection as described
in the methods section for the prediction of a negative outcome included PO (OR 1.05
per 1 %; 95 %-CI 1.02–1.09; P = 0.002), CRP value (OR 1.01; 95 %-CI 1.00–1.02; P = 0.027),
and RLC (OR 0.95; 95 %-CI 0.91–0.99; P = 0.011) ([Table 4]). The C-index for the model was 0.87 (95 %-CI: 0.80–0.95), indicating a good discrimination
between positive and negative outcome.
Table 4
Multivariable logistic regression model for the prediction of a negative patient outcome.
Tab. 4 Multivariates Modell zur Vorhersage eines negativen Outcomes.
|
Predictor
|
OR (95 %-CI)
|
p
|
c-index (95 %-CI)
|
|
PO (%)
|
1.05 (1.02–1.09)
|
0.002
|
0.87 (0.80–0.95)
|
|
CRP (mg/l)
|
1.01 (1.00–1.02)
|
0.027
|
|
RLC (%)
|
0.95 (0.91–0.99)
|
0.011
|
PO: percentage of opacity; OR: odds ratio; CI: confidence interval; c-index, concordance
index.
Based on these findings, we developed a score to facilitate the application of the
multivariable model in the clinical routine by dichotomizing each predictor. The cut-off
values were chosen at 39 % for PO, 80 mg/l for CRP and 15 % for RLC. A 4-point score
(0–3) was established: 1 point each if the cut-off for PO or CRP was exceeded or if
RLC fell below the cut-off. Based on the total score, the risk for a negative course
of disease in our cohort was as follows: 7 % for a score of 0, 30 % for a score of
1, 67 % for a score of 2, and 100 % for a score of 3 ([Table 5]).
Table 5
Predictive value of established score on outcome.
Tab. 5 Prognostische Wertigkeit des entwickelten Scores.
|
Cohort for model development (N = 83)
|
Validation cohort
(N = 28)
|
|
Score
|
Positive outcome
|
Negative outcome
|
Positive outcome
|
Negative outcome
|
|
0
|
25 (93 %)
|
2 (7 %)
|
17 (94 %)
|
1 (6 %)
|
|
1
|
14 (70 %)
|
6 (30 %)
|
1 (50 %)
|
1 (50 %)
|
|
2
|
7 (33 %)
|
14 (67 %)
|
0 (0 %)
|
4 (100 %)
|
|
3
|
0 (0 %)
|
15 (100 %)
|
0 (0 %)
|
4 (100 %)
|
Positive outcome: discharge or regular ward care; negative outcome: need for mechanical
ventilation, treatment in intensive care unit, ECMO, or death.
Internal validation
To validate the score, various methods were applied:
Predictive performance
In our dataset, the C-index (AUC) for the prediction of a negative outcome was high
(0.87). Somers' Dxy rank correlation between the predicted probabilities and the observed responses was
0.744.
Bootstrap validation
An internal validation of the score was performed to obtain stable optimism-corrected
estimates by using bootstrap validation with n = 1000 repetitions. The optimism corrected
c-index was 0.87, and the corrected Dxy was 0.744.
Calibration plot
A calibration plot comparing the predicted probabilities to the actual probabilities
is shown in [Fig. 2]. Both the apparent and bias-corrected probabilities are shown in comparison to the
ideal line. The score shows good overall performance.
Fig. 2 Calibration plot comparing the predicted probability of a negative outcome to the
actual probability.
Abb. 2 Kalibrierungsplot mit dem Vergleich der vorhergesagten und tatsächlichen Wahrscheinlichkeit
eines negativen Outcomes.
External validation
Applying the score to the validation cohort, the predictive performance was similar
to the model development cohort. A negative outcome was seen for 6 % of the patients
with a score of 0, 50 % for a score of 1, and 100 % for a score of 2 or 3 ([Table 5]).
Discussion
In this study, it was found that AI-based automatic lung segmentation and quantification
of opacities in COVID-19 pneumonia is feasible in most cases ([Fig. 3], [4]). All calculated values (LV, VO, PO, VHO, PHO, MHUL) correlated with patient outcome.
The percentage of opacified lung volume showed the best suitability for the prediction
of a bad outcome with an AUC of 0.83. The percentage of opacity and percentage of
high opacity demonstrated such a high degree of correlation that an evaluation stratified
according to the density of opacities did not add any additional information, i. e.,
separate investigation of ground glass opacity and consolidation was not helpful in
our cohort. With respect to laboratory values, a higher CRP, troponin, and LDH as
well as a lower RLC were significantly associated with a negative clinical outcome.
Multivariate regression analysis revealed a higher PO, higher CRP value, and lower
RLC to be the best predictors for an adverse outcome. To enhance the predictive power,
a model combining the aforementioned values was established. The C-index for the combined
model was exceptionally high (0.87). To facilitate integration of the combined model
into the clinical routine, a 4-point score was developed applying the cut-off values
of 39 % for PO, 80 mg/l for CRP, and 15 % for RLC. There is a great need for a prognostic
score tailored to patients suffering from COVID-19 pneumonia because it was shown
that an established score like the CURB-65 did not perform well [13]. The strength of the proposed score is the combination of a powerful imaging feature
with readily available laboratory data. All three components can be quickly and easily
determined in any hospital regardless of the existing infrastructure. The score was
able to separate patients into clearly defined risk groups ranging from 7 % for a
score of 0 to 100 % for a score of 3. The accuracy of the score was confirmed in a
validation cohort with excellent predictive performance.
Fig. 3 CT image of a 65-year-old patient with mild COVID-19 pneumonia and correct automatic
segmentation. The blue, light green, violet, and dark green lines show the segmentation
of the different lung lobes. Opacified lung areas are marked by a red line, and high
opacity is marked by a violet line.
Abb. 3 CT eines 65-jährigen Patienten mit einer milden COVID-19-Pneumonie und korrekter
Segmentierung. Die blaue, hellgrüne, violette und dunkelgrüne Linie markieren die
Segmentierung der Lungenlappen. Dichteangehobene Areale sind von einer roten Linie
markiert, Konsolidierungen von einer violetten Linie.
Fig. 4 CT image of a 39-year-old patient with severe COVID-19 pneumonia and correct automatic
segmentation. Color coding same as in Fig. 3.
Abb. 4 CT eines 39-jährigen Patienten mit schwerer COVID-19-Pneumonie und korrekter Segmentierung.
Farbige Markierungen wie in Abb. 3.
The correlation of the burden of opacifications on chest CT with clinical outcome
was demonstrated in earlier studies [8]
[9]
[10]
[13]. However, cut-off values were reported to be much lower compared to our results
(Colombi et al.: 27 %, Grodecki et al.: 1.8 % for consolidation and 13.5 % for ground
glass opacities, Lanza et al.: 23 %, Gieraerts et al.: 20 %). The most probable explanation
for this discrepancy is the different composition of the cohorts. CT scans in our
cohort were performed later in the course of disease (mean 9.6 days after symptom
onset in our cohort as opposed to 5–7 days in the cited studies). Furthermore, the
percentage of adverse outcomes was high in our cohort (42.7 %) compared to the other
studies (15.6–46 %). The possible link is the fact that one of the recruiting hospitals
in our study was a tertiary referral hospital equipped with an ECMO unit. This factor
results in a bias towards seriously diseased patients in our cohort. In line with
this circumstance, age was not a prognostic factor in our cohort as it has been in
most other investigations [25]
[26]. Taken together, the results of our investigation tend to include seriously rather
than to mildly diseased patients. This is not necessarily a disadvantage because a
viable prognostic score is especially needed for these at-risk patients.
There are some limitations to the study. The CT images were acquired in two different
hospitals and were heterogeneous regarding examination protocols. 21 of 89 CT scans
were contrast-enhanced for clinical reasons. The lung segmentation algorithm performed
equally well in both subgroups. Nevertheless, heterogenous examination protocols might
have affected the results of automatic image analysis and could have introduced a
bias. It has been shown earlier that results of lung densitometry are affected by
several technical parameters [28]. However, reconstruction algorithm, slice thickness, CT dose, CT scanner, and iterative
reconstruction have a minimal effect on mean lung density measurements [29]. The most important factor is the application of contrast media. Heussel et al.
showed that the median lung density increased by 18 HU after contrast application
in a cohort of patients with pulmonary emphysema [27]. The median lung density in our cohort differed substantially between the subgroups
(165 HU). Therefore, the difference created by contrast material seems to be negligible
in our cohort of patients with COVID-19 pneumonia.
Automatic lung segmentation was incorrect in 16 cases, especially in the case of CT
scans with large areas of subpleural consolidation or pneumothorax ([Fig. 5]). The version of the software prototype used in this study did not allow for manual
correction of lung segmentation so that these CTs had to be excluded from evaluation.
Meanwhile the software was further improved to allow manual corrections by the user.
Thus, exclusion of cases may no longer be necessary. Software-based lung analysis
was not tested for reproducibility. However, the group of Gieraerts et al. used the
same software and was able to show a high intrareader reproducibility [8].
Fig. 5 CT image of a 69-year-old patient with severe COVID-19 pneumonia and pneumothorax.
Erroneous lung segmentation most likely caused by pneumothorax. Color coding same
as in Fig. 3. This patient was excluded from the analysis.
Abb. 5 CT eines 69-jährigen Patienten mit schwerer COVID-19-Pneumonie und Pneumothorax.
Fehlerhafte Lungensegmentierung, am ehesten bedingt durch den Pneumothorax. Farbige
Markierungen wie in Abb. 3. Dieser Patient wurde von der Auswertung ausgeschlossen.
Further research is necessary to validate our findings in a larger validation cohort.
Furthermore, it remains to be assessed if our prognostic model could be enhanced by
integrating more parameters, like comorbidities or distribution of lung abnormalities
in the score.
In summary, in our cohort of patients with advanced COVID-19 pneumonia, the automatic
software-based determination of PO on chest CT images was a highly accurate prognostic
factor. The combination of PO with two readily available laboratory parameters (CRP
and RLC) further enhanced the prognostic power and made it possible to establish an
easy and quick to employ 4-point score for individual risk stratification.
-
A high PO in chest CT in COVID-19 pneumonia is a highly prognostic factor for a negative
outcome.
-
The combination of PO with two readily available laboratory parameters (CRP and RLC)
further enhances the prognostic power.
-
A 4-point scoring system based on these values allows quick and easy to employ individual
risk stratification.
Abbreviations
AI:
artificial intelligence
CK:
creatine kinase
CK-MB:
creatine kinase MB
COVID-19:
coronavirus disease 2019
CRP:
C-reactive protein
CTPA:
CT pneumonia analysis
ECMO:
extracorporeal membrane oxygenation
ICU:
intensive care unit
IL-6:
interleukin 6
LDH:
lactate dehydrogenase
LV:
lung volume
MHUL:
mean Hounsfield units of the whole lung
PCR:
polymerase chain reaction
PHO:
percentage of high opacity
PO:
percentage of opacity
REC:
relative eosinophil count
RLC:
relative lymphocyte count
VHO:
volume of high opacity
VO:
volume of opacity
WBC:
leukocyte count
