Free-Breathing Low-Field MRI of the Lungs Detects Functional Alterations Associated With Persistent Symptoms After COVID-19 Infection

With the COVID-19 pandemic, diagnostic imaging of the lungs has become crucial for the evaluation and follow-up of the alterations incurred by this infectious disease still to be fully characterized. Indeed, persistent symptoms have been observed more than 1 year after infection.1 Up to now, the main technology used for lung imaging in the clinical routine remains CT, which exposes the patient to ionizing radiation, especially with repetitive examinations, and has limited capability to assess lung functions. In contrast, MRI is a nonionizing technology providing functional information beyond morphologic assessment, but is hampered in the lungs by susceptibility effects due to the direct tissue-air interface,2 limiting its clinical use. However, these effects are mitigated at low field strengths, which are currently reconsidered by MRI vendors in conjunction with the latest advances in sequence developments, reconstruction, and data processing (eg, 0.55 T MRI).3,4 Indeed, promising whole-lung images were obtained at 0.55 T in COVID-19 patients, depicting the CT equivalent of ground-glass opacities.5,6 Moreover, Fourier-decomposition analysis techniques demonstrated the ability to assess lung functions without the need of breath-holds or exogenous contrast agent, providing defect maps correlating with gold standards such as dynamic contrast-enhanced MRI,7–12 hyperpolarized 3He MRI,8perfusion single photon emission CT,11129Xe ventilation MRI,13 and pulmonary function test (PFT) parameters.13 In particular, the so-called phase-resolved functional lung (PREFUL)14 implementation can provide a regional evaluation of perfusion and ventilation functions by means of quantitative maps, which enabled significant changes in defect areas to be identified in different pathologies.11,14–16 From a free-breathing acquisition of the lungs with a moderate temporal resolution (~300 milliseconds), the PREFUL processing technique reconstructs the full cardiac and respiration cycles after registration of the image time series and sorting of the time points based on a piecewise sine fit (for the cardiac cycle) and a cosine modeling (for the respiration cycle) of the average regional signal. As the signal intensity is related to the tissue proton density, which itself varies depending on the blood water content of the tissue and the lung volume, perfusion and ventilation parameters can then be extracted. The recent application of a modified version of PREFUL at 3 T in COVID-19 patients around 60 days after infection measured a larger ventilation volume with more heterogeneous and aggregated regions of high ventilation compared with healthy volunteers and increasing with the severity of infection.17

This exploratory study aimed to investigate the functional abnormalities that free-breathing lung MRI at 0.55 T in combination with the PREFUL processing technique could identify in patients with persistent symptoms after COVID-19 infection, as a potential application for the evaluation and follow-up of the disease.

MATERIALS AND METHODS Study Participants and Data Collection

This prospective study was performed after institutional review board approval with participants being consecutively included after written informed consent. Participants had to be older than 18 years and eligible for MRI examination.

First, 74 participants (further referred as patients) diagnosed with severe acute respiratory syndrome coronavirus 2 (COVID-19) determined by a positive reverse transcription–polymerase chain reaction test were consecutively included for this study. Five months on average after infection (from July to October 2020 in Germany), they were scanned according to the MR protocol described below on a prototype version of the 0.55 T whole-body Free.Max MRI system5 (Siemens Healthcare, Erlangen, Germany). In addition, at the time of acquisition, preexisting conditions affecting baseline lung functions (asthma or known reduced lung capacity) and persistent symptoms (if existent) were collected anamnestically by the responsible medical doctor. Despite not yet available at the time of the study, chosen anamnestic criteria were amongst spectrum of the later established World Health Organization consensus definition of post–COVID-19 condition.18 However, no physical medical examination was performed. The persistent symptoms encountered in this cohort were as follows: dyspnea (n = 16), cough (n = 1), respiratory/chest pain (n = 5), low cardiopulmonary reserve capacity (n = 1), dysgeusia (n = 9), dysosmia (n = 9), impaired concentration (n = 2) or stress tolerance (n = 1), reduced endurance or muscle power (n = 14), fatigue (n = 3), palpitations (n = 2) or sweating (n = 2), intensified preexisting allergy (n = 1) or asthma (n = 1), vision problems (n = 3), word-finding (n = 1) or swallowing (n = 1) disorders, dizziness (n = 1), gastritis (n = 1), sensation of cold (n = 1), cold feet (n = 1), capitis (n = 1), hair loss (n = 3), and cervical lymph nodes swelling (n = 1). Eighteen patients had more than one of the aforementioned symptoms. Thirteen patients were hospitalized, but none were intubated.

To collect normative values, 8 additional participants without known history of COVID-19 infection (further referred as healthy volunteers) were included and scanned with the same MRI protocol on the clinical 0.55 T Free.Max system.

Demographics and statistics of the cohort are reported in Table 1.

TABLE 1 - Demographics and Statistics of the Final Cohort Sample Size Sex Age, y Weight, kg Height, cm BMI, kg/m2 Time From Infection, d Patients without persistent symptoms 28 7 women, 21 men 44.6 ± 17.2 [21–78] 77.4 ± 14.0 [47–113] 176 ± 9 [158–196] 24.9 ± 3.8 [18.8–33.9] 161.5 ± 31.0 [106–225] Patients with persistent symptoms 42 17 women, 25 men 44.7 ± 12.2 [24–67] 80.8 ± 16.9 [53–148] 177 ± 8 [161–194] 25.8 ± 4.9 [19.4–47.8] 164.3 ± 24.9 [114–227] Total patient cohort 70 24 women, 46 men 44.6 ± 14.3 [21–78] 79.4 ± 15.8 [47–148] 176 ± 8 [158–196] 25.5 ± 4.5 [18.8–47.8] 163.2 ± 27.3 [106–227] Healthy volunteers 8 4 women, 4 men 33.5 ± 7.8 [25–45] 69.9 ± 11.2 [55–85] 173 ± 10 [157–185] 23.1 ± 1.9 [20.7–25.9]

Note: 1 patient aborted the acquisition because of claustrophobia, and 3 patients had to be discarded after data processing because the registration algorithm failed to converge, resulting in a final patient cohort of 70 patients. Statistics are indicated as mean ± standard deviation [minimum-–maximum]. The average body mass index (BMI) was at the upper limit of the standard range.

One patient aborted the acquisition because of claustrophobia, and 3 patients had to be discarded because the registration algorithm failed to converge during data processing, resulting in a final cohort of 70 patients and 8 healthy volunteers. Although the final cohort of patients included 1.5 times more patients with persistent symptoms (n = 42) than without (n = 28), age, weight, height, body mass index (BMI), and time between measurements and infection were similar in the 2 groups.

MR Data Acquisition

A balanced steady-state free-precession (bSSFP) sequence (TrueFISP) was set up for free-breathing lung imaging and PREFUL postprocessing. The final parameters were as follows: one 2-dimensional central coronal slice positioned at the middle of the hila, thickness = 15 mm, in-plane resolution = 1.7 × 1.7 mm2, matrix = 128 × 128 (interpolated to 256 × 256), bandwidth = 1302 Hz/pixel, flip angle = 30 degrees, TR/TE = 276.7/1.6 milliseconds, GRAPPA = 2, no partial Fourier, 250 time points, temporal resolution = 376.7 milliseconds, duration = 1 minute 34 seconds.

Two-dimensional turbo-spin-echo (TSE) transverse and coronal morphological sequences with BLADE (periodically rotated overlapping parallel lines with enhanced reconstruction) readout and respiration-gated with a navigator were added to the protocol. The transverse image was proton density-weighted (TE/TR = 34/3000 milliseconds) with 1.1 × 1.1 mm2 in-plane resolution, 336 × 336 matrix, thirty 6 mm-thick slices. The coronal image was acquired with a STIR preparation (short-tau inversion recovery), T2-weighting (TE/TR = 74/2500 milliseconds), 1.5 × 1.5 mm2 in-plane resolution, 272 × 272 matrix, thirty 6 mm-thick slices.

Data Processing Functional Parameters Calculation

The following metrics were automatically calculated voxel-wise with a prototype implementation of the PREFUL technique14 (MR Lung v2.0; Siemens Healthcare, Erlangen, Germany), after registration to a midexpiration position, which was manually selected in the middle of the acquisition based on the respiratory trace:

Normalized perfusion (Q, %)11; Fractional ventilation (FV, %)19; Flow-volume loop correlation (FVLc)16; Perfusion time-to-peak (qTTP, milliseconds), also referred as pulmonary pulse wave transit time; and Absolute deviation of the FV time-to-peak from the peak inspiration (vTTP, % of the respiratory cycle).

Each of these parameters is defined in more details in Table 2.

TABLE 2 - Definition of the PREFUL Parameters and Calculation Methods Notation, Units Definition Q, % Perfusion normalized with respect to a full-blood signal region, derived as follows: (1) find the parenchyma cardiac cycle phase by histogram analysis of the signal in the lung parenchyma, determined as the time point in the reconstructed cardiac cycle when most of the lung parenchyma voxels reach their maximal value; (2) determine the full-blood signal region for normalization as the highest maximal intensity projection value in between the lungs and expected to reflect the aorta or other available large vessel11; and (3) divide the map of the reconstructed cardiac cycle signal at the parenchyma phase by the maximal value in the normalization region. FV, % Fractional ventilation calculated as:

SmidSinsp−SmidSexp

with S the signal value at end-inspiration (insp), end-expiration (exp) and middle position (mid)19 FVLc, −1 to 1 Flow-volume loop correlation16: correlation of the flow-volume loop (deduced from the FV reconstructed cycle*) with respect to a healthy region (largest connected region within the 80th and 90th FV percentiles) qTTP, ms Perfusion time-to-peak with respect to the normalization region (see the definition of Q), also referred as pulmonary pulse wave transit time vTTP, % Absolute deviation of the FV time-to-peak from the peak inspiration, expressed in percentage of the respiratory cycle Q-Defect-Total, % Percentage of areas identified as Q defect (values below 2%) FV-Defect-Total, % Percentage of areas identified as FV defects (values below 40% of the 90th percentile in the lungs) FVLc-Defect-Total, % Percentage of areas identified as FVLc defects (values below 0.9) Q-FV-Defect, % Percentage of areas with concurrent Q and FV defects Q-FVLc-Defect, % Percentage of areas with concurrent Q and FVLc defects Q-Defect-FV-Exclusive, % Percentage of areas with Q defect but without FV defect Q-Defect-FVLc-Exclusive, % Percentage of areas with Q defect but without FVLc defect FV-Defect-Q-Exclusive, % Percentage of areas with FV defect but without Q defect FVLc-Defect-Q-Exclusive, % Percentage of areas with FVLc defect but without Q defect Q-FV-Non-Defect, % Percentage of areas without Q defect and without FV defect Q-FVLc-Non-Defect, % Percentage of areas without Q defect and without FVLc defect

*Note: the FV reconstructed cycle is the fractional ventilation value along the respiration cycle, which can be reconstructed voxel-wise with the PREFUL technique (sorting of the data points along the inspiration and expiration phase followed by an interpolation on an equidistant time grid).


Functional Defect Percentage Calculation

To calculate the percentage of Q, FV, and FVLc defects, thresholds below which values in the calculated maps should be considered as abnormal had to be defined. These thresholds were determined after a similar methodology and on the same dataset as described in Behrendt et al.12 In more details, this dataset consisted in 2 to 10 coronal slices acquired in 62 healthy volunteers and 87 patients with various lung diseases (23 cystic fibroses, 43 chronic obstructive pulmonary diseases, 21 chronic thromboembolic pulmonary hypertension) at 1.5 T with a spoiled gradient-echo sequence. The following thresholds were considered:

For Q: Q values from 0% to 3% with 0.25% increment (13 different thresholds); For FV: the 50th, 75th, 90th, and 97th percentile values in the lung parenchyma multiplied with factors from 0.1 to 0.9 with 0.1 increment (45 different thresholds); and For FVLc: FVLc values from 0.1 to 0.99 with 0.05 increment (18 different thresholds)

For each threshold and functional parameter (Q, FV, FVLc), the following steps were performed:

For every healthy volunteer and patient, the percentage of defects was calculated. The receiver operating characteristic (ROC) curve was computed for the entire dataset, and the area under the curve (AUC) and optimal operating point were determined. Based on the optimal operating point, all subjects were relabeled as healthy volunteer or patient. Comparing these new labels with the ground truth, the confusion matrix was calculated and the accuracy, precision, sensitivity, specificity, and F-score of the considered threshold were derived.

The selection criteria for the optimal threshold were as follows:

High accuracy, sensitivity, specificity, precision, F-score, and AUC; Low false-positive and false-negative rates; A median defect percentage below 5% in healthy volunteers; and A large difference in defect percentage between healthy volunteers and patients.

The thresholds of 2% for Q, 40% of the 90th percentile value for FV and 0.9 for FVLc were the thresholds optimizing the trade-off between those criteria.

Applying those thresholds to the maps calculated at the previous section, the percentage of defect areas (Q-Defect-Total, FV-Defect-Total, FVLc-Defect-Total) were calculated. The percentage of concurrent defect areas of perfusion and ventilation metrics (Q-FV-Defect, Q-FVLc-Defect) and perfusion defects exclusive to ventilation defects (Q-Defect-FV-Exclusive, Q-Defect-FVLc-Exclusive) were then derived, and vice versa (FV-Defect-Q-Exclusive, FVLc-Defect-Q-Exclusive), in addition to areas without defect on both perfusion and ventilation maps (Q-FV-Non-Defect, Q-FVLc-Non-Defect). These defect parameters are also individually defined in more details in Table 2.

Data Analysis Variable Selection for the Detection of Persistent Symptoms

To assess the redundant information conveyed by all MRI metrics described in Table 2, the Pearson correlation matrix was calculated with a significance level of P value <0.05.

Then, to investigate the functional correlates of persistent symptoms, the variables with the highest detection power for the presence of persistent symptoms were determined. All MRI metrics, as well as sex, age, BMI, and the presence of preexisting conditions, were included.

The model of choice was therefore a logistic regression, expressed as pv=11+e−∑i=1Nαivi with p(v) the probability to belong to the group of patients with persistent symptoms, v = (v1, …, vN) the variable values of the patient, and αi the associated regression coefficients.

The variable values were first standardized with regards to their respective distribution in the cohort. Then, for every number of variables N to select, the recursive feature elimination algorithm20 (as implemented in scikit-learn 1.0.121) was applied with a logistic regression model for the detection of persistent symptoms through a Repeated Stratified K-Fold cross-validation strategy with 7 splits (10 patients per fold, 6 folds for fitting, 1 for evaluation), repeated 5 times. In each fold, the ratio between the numbers of patients with and without persistent symptoms in the entire dataset was approximately respected. The procedure is illustrated in Figure 1. In more detail, the recursive feature elimination algorithm recursively eliminated the variables with the least importance (determined by the square of the fitted regression coefficient) based on the training set of each split, to keep the N most informative variables. Then, the logistic regression model was fitted with the N selected variables on the same training set and was applied to the test set to calculate the detection accuracy (fraction of correctly classified patients). This cross-validation strategy therefore resulted in 35 accuracy scores (7 different splits with 1 test fold each, 5 repetitions) for each number N of selected variables.

F1FIGURE 1:

Strategy to determine the most relevant variables for the detection of persistent symptoms. A repeated stratified k-fold cross-validation strategy was used, with 7 folds (10 patients/fold) and 5 repetitions, generating 5 times 7 splits with 6 folds used for training (orange rectangles) and 1 fold for testing (green rectangles). Note that the class ratio of the entire dataset was approximately respected for each fold. The considered variables were all the extracted functional MRI metrics, sex, age, body mass index, and the presence of preexisting conditions.

Detection Accuracy of the Selected Variables

The set of variables providing the highest mean accuracy was kept. The mean ROC curve was calculated with a Stratified K-Fold cross-validation (7 splits). Finally, the logistic regression model was fitted on the entire dataset. The score λ=∑i=1Nαiνi, with αi the fitted coefficient for the variable νi, was computed for each patient, and a 2-sided Student t test was performed between patients with and without persistent symptoms.

For comparison, the variable selection process was repeated excluding all MRI metrics. In addition, 2-sided Student t tests and proportion-based z-test (for sex and existence of preexisting conditions) between patients with and without persistent symptoms were performed for each variable (Statsmodels 0.12.222).

The MR data of the healthy volunteers, all codes, and necessary data to reproduce this analysis and the resulting figures are publicly available at github.com/slevyrosetti/lung-MRI.

Review of the Morphological Images

For comparison purposes, the morphological images of patients and healthy volunteers were evaluated and scored by a board-certified radiologist with more than 5 years of experience who was blinded to the subjects' clinical presentation and functional data. The scoring method previously elaborated for follow-up CT examinations of COVID-19 patients23 was used. More precisely, the following imaging patterns were listed for each of the 5 lobes (upper right lobe, middle lobe, lower right lobe, upper left lobe, lower left lobe) according to the Fleischner Society glossary24: ground-glass opacities, consolidation, reticulation, emphysema, pleura thickening, pleural effusion, nodules or masses, honeycombing, bronchiectasis, and interlobar pleural traction. The extent of abnormalities was then quantified in each lobe using the 6-point scoring system23: 0 for no involvement, 1 for less than 5% involvement, 2 for 5% to 25% involvement, 3 for 26% to 49% involvement, 4 for 50% to 75% involvement, and 5 for more than 75% involvement. Finally, the total abnormality score was obtained by summing up the score of each lobe, yielding a score between 0 and 25.

The Pearson correlation of the total morphological abnormality score with each of the functional parameters extracted previously was calculated. A 2-sided Student t test between patients with and without persistent symptoms was also performed.

RESULTS

The quality of the image time series obtained in 2 patients, without and with persistent symptoms, can be observed in the Video Supplemental Digital Content 1, https://links.lww.com/RLI/A711. Pulmonary branches could be well visualized and the banding artifacts did not compromise the image quality in the region of interest.

High significant (P < 0.05) correlations were observed between functional MRI metrics (Fig. 2), highlighting the information redundancy and, consequently, the necessity to carefully select the most relevant variables for the detection of persistent symptoms. For example, Pearson correlation coefficient of Q-Defect-FVLc-Exclusive was 1.0 with Q-Defect-FV-Exclusive and 0.99 with Q-Defect-Total. Generally, metrics related to FV and FVLc were highly correlated (Q-FV-Non-Defect and Q-FVLc-Non-Defect correlated at 0.96, Q-FV-Defect and Q-FVLc-Defect correlated at 0.95). In contrast, the abnormality score derived from the morphological images review showed poor correlation (magnitude ≤0.37) with the functional metrics extracted with PREFUL.

F2FIGURE 2:

Pearson correlation coefficient between MRI metrics. Insignificant correlations (P > 0.05) were shaded. Large redundancy between multiple MRI metrics can be observed, with correlation up to 1.0, hence the necessity to carefully eliminate redundant variables. The morphological abnormality score (morpho. score) derived from the review of the morphological images showed poor correlation with functional MRI parameters extracted with PREFUL.

Results of the variable selection algorithm (Fig. 3) revealed that the highest accuracy (average of 66.9%, median of 70.0% across splits and repetitions) in the detection of persistent symptoms was obtained with 2 variables. The 2 most relevant variables were:

F3FIGURE 3:

Results of the variable selection for the detection of persistent symptoms. The x-axis lists the variables by order of selection. As the newly selected variable can vary across splits and repetitions, the indicated variable is the most often selected variable over all splits and repetitions and the occurrence of the selection is indicated by the bottom orange bar chart as a measure of stability. The red line (right y-axis units in red) represents the mean accuracy across splits of the logistic regression model tested with the corresponding number of selected variables. The blue bar chart (left y-axis units in blue) represents the mean relative importance across splits of the newly selected variable (squared coefficient of the variable with respect to the total sum of squared coefficients) indicated on the x-axis when the model was tested with the corresponding number of variables. The maximum mean and median accuracy obtained were 66.9% and 70.0% using the variables Q-FVLc-Defect and Q-FVLc-Non-Defect, respectively. Including more variables likely added more noise than relevant information for the detection of persistent symptoms.

Q-FVLc-Defect, and Q-FVLc-Non-Defect.

Because of information redundancy, including more variables likely added more noise than relevant information since the accuracy dropped to 60.6% when including 3 variables. The relative importance of Q-FVLc-Defect and Q-FVLc-Non-Defect were 67.8% and 33.1%, respectively (average across splits and repetitions), when evaluating the model with 2 variables. Moreover, the selection of these 2 variables was stable across splits and repetitions, with Q-FVLc-Defect and Q-FVLc-Non-Defect selected in 97% and 86% of the times, respectively. For comparison, when discarding MRI metrics, the highest detection accuracy obtained was 57% with one variable (sex).

The mean AUC when fitting and testing the model with the selected variables on different datasets was 0.68 (Fig. 4). When fitting the model on the entire dataset, the detection accuracy was 71.4% and the AUC was 0.71, with the respective coefficients as follows:

F4FIGURE 4:

Receiver operating characteristics (ROC) curve analysis with a stratified k-fold cross-validation using 7 splits (10 patients/fold, 6 folds for fitting, 1 for testing, with approximately identical class ratios in each fold as in the entire dataset), based on the Q-FVLc-Defect and Q-FVLc-Non-Defect values. The blue bold line is the mean ROC curve across the 7 individual splits, which are also represented in thinner colored lines. The gray-shaded area delineates ±1 standard-deviation around the mean ROC curve. The dotted curve was derived from the fit of the logistic regression model on the entire dataset (same data for training and testing, hence the higher AUC). The red dashed line represents the random classifier.

α1 = 3.19 (without standardization, α1native = 1.02) α2 = 1.59 (α2native = 0.09)

Although no significant difference in Q-FVLc-Defect and Q-FVLc-Non-Defect were observable between patients with and without persistent symptoms (P value of 0.233 and 0.227, respectively), the difference in the score λ = α1(native) × [Q-FVLc-Defect] + α2(native) × [Q-FVLc-Non-Defect] was significant (P = 0.017).

The median value and interquartile range of each parameter for each group as well as the P value of the statistical tests between patients with and without persistent symptoms were reported in Table 3. When available in the literature,13,15,25–27 mean healthy values across studies at 1.5 T were added for comparison. All defect (or nondefect) percentage parameters showed a consistent increase or decrease from healthy volunteers to patients without persistent symptoms to patients with persistent symptoms. In particular, the score λ increased from 7.7 to 8.2 to 8.6. In contrast, functional parameters averaged across the lungs (mean perfusion, mean FV) did not show a consistent trend, demonstrating the increased sensitivity and added value provided by regional parameters over lung-averaged parameters. As expected, the mean morphological abnormality score showed a poor discrimination between the groups (0.5 vs 0.6 vs 1.3), with a median value of 0 for all groups. Regarding the comparison of absolute values with 1.5 T spoiled gradient-echo data in healthy subjects, large differences, in particular in ventilation defects (4.7-fold difference for FVLc-Defect-Total), were measured using the same thresholds, highlighting the necessity to determine new threshold values for 0.55 T bSSFP data.

TABLE 3 - Median [First Quartile–Third Quartile] Values of All Variables Compared With Healthy Values Available in the Literature and P Values of Statistical Difference Between Patients With and Without Persistent Symptoms After COVID-19 Infection Variable Mean Healthy Values at 1.5 T (Literature13,15,25–27) Healthy Values at 0.55 T (Current Study, n = 8) Without Persistent Symptoms (n = 28) With Persistent Symptoms (n = 42) P Sex 50.0% female, 50.0% male 50% female, 50% male 25.0% female, 75.0% male 40.5% female, 59.5% male 0.18 Age, y 28.4 33.5 [26.5–38.2] 42.5 [30.6–60.2] 41.4 [35.6–52.6] 0.98 Preexisting conditions occurrence 12.5% 14.3% 11.9% 0.77 BMI, kg/m2 22.3 22.8 [21.8–24.8] 24.4 [22.8–26.6] 25.2 [22.6–27.9] 0.43 Q-FVLc-Defect, % 0.3 [0.1–1.2] 0.3 [0.0–1.0] 0.2 [0.0–1.8] 0.23 Q-FVLc-Non-Defect, % 69.3 [62.3–82.4] 82.1 [70.2–88.7] 87.0 [78.3–92.5] 0.23 Mean FVLc 0.96 0.90 [0.88–0.91] 0.95 [0.93–0.97] 0.96 [0.94–0.97] 0.21 Mean perfusion, % 6.0 9.3 [7.4–12.8] 8.9 [6.5–14.3] 9.2 [6.1–12.1] 0.87 Mean FV, % 14.1 12.8 [9.1–15.8] 11.6 [9.6–16.3] 14.4 [10.6–20.1] 0.15 Q-Defect-FVLc-Exclusive, % 1.4 [0.6–7.1] 2.6 [1.5–17.0] 3.3 [2.0–9.1] 0.32 Q-FV-Non-Defect, % 77.3 [73.5–80.3] 83.0 [74.9–88.0] 84.9 [78.2–90.1] 0.39 Q-Defect-FV-Exclusive, % 1.6 [0.5–6.5] 3.0 [1.6–17.0] 4.1 [1.9–9.8] 0.41 FVLc-Defect-Q-Exclusive, % 21.0 [14.4–33.4] 9.6 [3.9–12.9] 6.3 [3.5–10.7] 0.10 FV-Defect-Q-Exclusive, % 20.1 [15.3–22.2] 8.4 [5.9–12.0] 8.4 [5.4–11.6] 0.52 FVLc-Defect-Total, % 6.0 23.0 [14.6–34.4] 10.0 [4.6–13.7] 7.7 [4.0–12.1] 0.40 Q-FV-Defect, % 0.2 [0.1–1.7] 0.2 [0.0–0.7] 0.3 [0.1–1.0] 0.45 FV-Defect-Total, % 5.5 20.1 [15.6–22.4] 9.8 [6.4–12.9] 8.7 [6.4–12.4] 0.78 Q-Defect-Total, % 5.9 2.2 [0.7–7.5] 4.0 [1.9–17.5] 4.6 [2.2–12.0] 0.50 qTTP, ms 95 65 [56–138] 91 [53–357] 120 [82–341] 0.86 vTTP, % 9.2 [6.6–11.6] 5.9 [4.1–7.9] 5.3 [4.5–7.0] 0.40 λ (without standardization) 7.7 [6.8–7.9] 8.2 [7.4–8.7] 8.6 [8.2–9.0] 0.02* Morphological abnormality score (0–25) 0.5 [0–1] 0.6 [0–1] 1.3 [0–1] 0.19 Note: cells were left empty when no healthy value was available in the literature. Sex and age healthy values are the mean values across all included studies13,15,25–27 and might not be representative for variables where not all studies reported a variable value. The reported P value is the P value of a 2-sided proportion z-test between the 2 patient groups in the case of categorical variables (sex, preexisting conditions occurrence) and a 2-sided Student t test in the case of continuous variables (all others). The score λ is defined as λ = α1native × [Q-FVLc-Defect] + α2native × [Q-FVLc-Non-Defect] with α1native and α2native the fitted coefficients without prior standardization of the [Q-FVLc-Defect] and [Q-FVLc-Defect] values. For the morphological abnormality score, the mean value [first quartile–third quartile] was reported as it was more informative than the median (0 in all groups).

*Significant difference in the variable value between patients with and without persistent symptoms at the significant level of 0.05.

Individual functional maps obtained in representative patients with and without persistent symptoms, with different classification probabilities, are presented in Figure 5. Patients wrongly classified by the model in each group are also shown. When comparing the 2 patients classified with the highest probability in each group (third and fourth columns), clear perfusion and ventilation abnormalities could be observed on the perfusion, FV, and FVLc maps of the patient with persistent symptoms, resulting in a larger proportion of concurrent perfusion and FVLc defects (Q-FVLc-Defect). These differences in lung fun

留言 (0)

沒有登入
gif