We conducted a retrospective study including consecutive patients with a clinical indication to undergo CMR who received an extended scan protocol including the PREFUL sequence. Patients scanned between September 2023 and January 2024 on a 1.5T scanner (Siemens Healthineers, AvantoFIT, Forchheim, Germany) were screened. Inclusion criteria were age > 18 years, available PREFUL image acquisition, and cine imaging in a short-axis stack covering the left and right ventricle (RV). Basic cohort characteristics such as age, height, weight, blood pressure and heart rate were acquired prior to the scan. Comorbidities were derived from the electronic patient record.

PREFUL MRI was performed with a fast low angle shot sequence with the following parameters: repetition time: 2.1 ms, echo time: 0.93 ms, field of view: 500 × 500 mm2, matrix: 128 × 128 (interpolated to 256 × 256), slice thickness: 15 mm, flip angle: 5°, parallel imaging acceleration factor: 2, temporal resolution: <300 ms/frame, number of measurements: 250. Total acquisition duration was 60 s per slice.

All cine images were based on balanced steady state free precession sequences with retrospective ECG gating. The short axis had the following scan parameters: repetition time: 2.78–3.31 ms, number of reconstructed phases: 30, echo time: 1.19–1.44 ms, field of view: 340–380 × 276–380 mm2, matrix: 192 × 156, voxel size: 1.8-2.0 × 1.8-2.0 mm2, slice thickness: 7 mm with no gap, flip angle: 74–80°, parallel imaging acceleration factor: 2.

Images were acquired in three slices with the middle one being positioned on the bifurcation of the pulmonary artery on standard localizer images (Fig. 1). Post-processing was carried out with a research software tool provided by the vendor (MR Lung v2.3.0; Siemens Healthineers, Forchheim, Germany). Details were published previously [15]. Briefly, ventilation- and perfusion-related signal changes were separated following image registration to an intermediate lung volume by applying low- and high-pass temporal filtering. After estimation of respiratory and cardiac phases, synthesized cycles with increased temporal resolution were reconstructed. Expiratory and inspiratory phases were used to compute regional ventilation (RVent) maps in ml/ml. Cardiac phase information related to parenchymal perfusion was obtained by histogram analysis and used to derive normalized perfusion (QN, expressed as percent) using the signal intensity of a full-blood voxel (e.g., aortic signal) as reference. For further analysis, a convolutional neural network was used to segment the lung parenchyma as the region of interest. Perfusion defects were identified using a fixed threshold of QN < 2.0%, while ventilation defects were determined using an adaptive threshold defined as 0.4 × 90th percentile of the RVent values within the lung parenchyma. The final output included total perfusion and ventilation defect percentages (QDP and VDP, respectively) for each slice as well as a weighted average of all three slices. A weighted mean across slices was calculated based on the lung area covered by each slice. Reproducibility parameters for the PREFUL sequence have been reported previously [16].

Slice planning for the PREFUL sequence. Slice planning for the PREFUL sequence (left) with the tracheal bifurcation used as the first reference acquisition and with adjacent posterior and anterior slices (maps shown on the right) for representative sampling of the whole lung

All image analyses were carried out by one reader (J.H.) with three years of CMR experience supervised by a second reader (J.G.) with 6 years of CMR experience. Post-processing for cardiac function was carried out using dedicated software (CVI42, version 5.13.7, Circle Cardiovascular Imaging, Calgary, Canada). Biventricular function and volume assessment were carried out in the short-axis cine stack. Endo- and epicardial contours were drawn for the LV in the end-diastolic phase, while only endocardial contours were drawn in end-systole. Papillary muscles were separately contoured in end-systole and end-diastole and included in the overall LV mass. The RV was contoured in end-diastole and end-systole (endocardial contours). All segmentations were carried out according to current recommendations [17].

To evaluate the feasibility of the PREFUL approach in clinical routine, we assessed both acquisition success and the achievement of diagnostic image quality. Additionally, during post-processing, we examined the accuracy of automatic segmentations and determined whether the quantitative outputs were interpretable.

Data is represented as median and interquartile range (IQR) for continuous variables and as number and percent for categorical variables. Normal distribution was assessed with the Shapiro-Wilk test. Continuous variables were compared globally with the Kruskal-Wallis test and in case of significance followed by pairwise comparisons with the Mann-Whitney U test with Bonferroni correction for multiple testing. Categorical variables were compared with the Chi-square test or Fisher’s exact test. Correlation analysis was carried out using Spearman’s correlation coefficient. In addition, model selection was performed by stepwise ordinary least squares regression with backward elimination of age, sex, body mass index (BMI), LV ejection fraction (LVEF), pulmonary disease (present/absent) as predictors for total QDP and total VDP using an R2 criterion. To correct for bias introduced by backward selection we performed model evaluation using a nonparametric bootstrap approach: Backward selection was repeated on 500 bootstrapped samples of the dataset, and for each variable retained in the final models, the mean coefficient, standard deviation, 95% confidence intervals, and inclusion frequency were calculated to provide a robust estimate of model parameters and prevent overfitting. Two-tailed empirical p-values were calculated at a level of p < 0.05. A p < 0.05 was defined as a statistically significant difference. Statistical tests were performed using IBM SPSS Statistics (version 29.0.0.0.) and Python 3.12.11, using the pandas 2.2, statsmodels 0.14.4 and scikit-learn 1.7.0 packages.