Two inclusion criteria were defined: patients should be ≥18 years old and with at least 1 non-calcified pulmonary nodule with a long axis ≥4 mm. The exclusion criteria are gathered in Table 1.

Eighty-two patients were identified as presenting the inclusion criteria. Seventy (85%) patients were not eligible for various causes (Figure 1), 1 patient agreed to participate but could not due to personal reasons.

The HF-NIV was performed on patients with a Monsoon III ventilator (Acutronic Medical Systems, Hirzel, Switzerland) and a non-invasive patient interface (Phasitron, Percussionnaire, Idaho, USA) (14), according to a set-up described in Delacoste et al. (9). The oxygen delivery has been set to 100% at the monitor. In practice, given the ingress of ambient air through the inspiratory port of the Phasitron ®, the fraction of inspired oxygen (FiO₂) was ~50% (10).

Before the MRI exam, to assess the tolerance to the HF-NIV and to ensure a respiratory stabilization period ≥6 min at a respiratory rate of 250/min in agreements with Ogna et al. (15), each patient was trained by a physiotherapist during a screening session. The following physiological parameters were also monitored by a pulmonologist with a Digital Monitoring System (SenTec, Therwil, Switzerland): continuous transcutaneous capnography (TcCO₂), oxygen saturation (SpO₂) and cardiac frequency. Arterial pressure was also monitored. Eligible patients underwent the MRI exam less than a month after the CT exam and due to MR constraints only SpO₂ and cardiac frequency were then controlled.

All data acquisitions were performed on a 3T clinical MR scanner (MAGNETOM PrismaFit, software version VE11D, Siemens Healthcare, Erlangen, Germany). All patients were positioned feet first, supine, with arms at their side. A 16-channel body array combined to 32-channel spine coils were used. Shimming was performed ahead of the acquisitions to correct for the magnetic field inhomogeneities.

A prototype double echo Ultra-Short Echo time (UTE) sequence (16, 17) with spiral phyllotaxis trajectory (18) was acquired without any triggering under HF-NIV (UTE-HF-NIV) and in free-breathing (UTE-FB), with the following parameters: echo times TE₁/TE₂ = 0.08 ms/2.86 ms, repetition time TR = 5.9 ms, readout bandwidths BWr₁/ BWr₂ = 305 Hz/Px/610 Hz/Px, radio-frequency excitation angle = 5°, field of view (FOV) = (250 mm)³, voxel size = 1 mm³, 1,220 segments of 50 readouts each and an acquisition time of 6 min. The UTE-FB at tidal volume was always acquired before the HF-NIV acquisition at full inspiration to avoid any change in breathing related to the potential recovery after the HF-NIV procedure.

The reconstruction was performed off-line with Matlab (Version R2018b, The Mathworks, Natick, Massachusetts, USA) and a non-uniform fast Fourier transform. For FB data, a compressed sensing algorithm that exploited the sparsity over the respiratory dimension was used to obtain respiratory-motion-resolved images. Only end-expiration images were kept for analysis. If the patient happened to breathe before the end of the HF-NIV acquisition, the motion-corrupted data were removed from the raw data set prior to the off-line reconstruction.

A VIBE sequence with asymmetric sampling in the readout dimension was also acquired during an unassisted breath-hold (BH) on inspiration (VIBE-BH). The acquisition parameters were as follows: TE = 0.94 ms, TR = 4.0 ms, BWr = 500 Hz/Px, radio-frequency excitation angle = 5°, FOV = 400 × 262.4 mm², voxel size = 1 × 1 × 3 mm³, slice partial Fourier = 7/8, orientation = transversal, CAIPIRINHA acceleration factor = 4, T_acq = 20 s. Images were directly reconstructed at the MR scanner.

The gold standard was a CT exam based on the ALARA principle and performed on a Revolution or Discovery scanner (General Electric, Milwaukee, USA, n = 7) or SOMATOM Force scanner (Siemens Healthcare, Erlangen, Germany, n = 1) (19). A low-dose and without intravenous contrast chest CT with a CTDI below 3 mGy [0.5–1.99], corresponding to a dose below 3 mSv, as recommended by the Fleichner Society and ACR recommendations was applied in all but one patient (20). The latter had a CT in the setting of oncological context with IV contrast administration and normal radiation dose. The mean of low dose CTs corresponded to an effective dose of 0.61 mSv. The normal dose CT performed in the setting of oncological follow-up corresponded to an effective dose of 3.32 mSv. Thin slices, with a slice thickness 1.25 mm, FOV 320 mm with in plane voxel size 0.625 mm, were used with soft and lung kernel reconstructions.

MR nodule's detection analysis was carried out by three experienced chest radiologists (reader 1 (R1) >20 years, R2, 8 years and R3, 4 years of experience in thoracic imaging) on anonymized and randomly sorted images. MR volume assessment was performed by the most experienced reader of the three (R1). Anonymized gold standard CT images were analyzed by two readers (R1 and R4, 5 years of experience in thoracic imaging). To avoid any bias, the common reader (R1) first performed the nodule's detection on MR images and after a 2 weeks period performed the reading of gold standard CT images.

The nodule detection was determined according to the following visual score: absent = 1; probably absent = 2; uncertain = 3; probably present = 4; present = 5. The true positive, false negative and false positive nodules were reported for each technique and each reader. The detection rate of each technique was assessed for nodules ≥4 mm and ≥6 mm by computing the fraction of nodules with a visual score >3, i.e., considered as present, in MR images and confirmed on CT. The two MR techniques that provided the highest detection rate were used for further analyses.

Evaluated morphological criteria included the long axis (l_axis), the mean diameter (d_mean, average of the long and short axes), the solid/subsolid texture and the intrapulmonary lymph node nature or not of solid nodules. The distance to the pleura (d_pleura) was assessed and nodules were divided within superior, middle or inferior areas for nodules in segments 1 to 3, 4 to 6, and 7 to 10, respectively, according to the Boyden classification.

The detection's dependency on the nodules' d_pleura and superior-middle-inferior location was investigated as well as on the l_axis, d_mean and measured volume in soft kernel images (V_soft).

The 2D-l_axis measurements were compared between the CT images and images of the two MR techniques with the highest detection rate of MR detected nodules. 3D measurements were performed on CT, and on maximum five randomly selected nodules per patient among all included nodules for MRI, this to minimize bias. When a selected nodule was not detected with one of the two investigated MR techniques, another nodule was selected, when available.

The nodules' volume was measured with the lesion management software (Carestream, Rochester, New York, USA), both on lung and soft kernel CT images by one reader, and soft kernel CT images only by a second reader. The volumetric measurements performed by the more experienced reader in soft kernel images were considered the reference value. Semi-automatic contouring was used by default for CT and manual drawing was used for MR and when necessary for CT. 3D volumetric measurements were compared between lung and soft kernel CT images and MR images.

The formula of Beam et al. was used to determine the minimal sample size of nodules required to demonstrate an improvement in sensitivity of 90% with a power of 80% (21). A two-sided Fisher's exact test was applied to investigate the dependency between the nodules' detection and their location in superior, mid, or inferior lung areas. To investigate the nodules' detection dependency on d_pleura, l_axis, d_mean, and V_soft, Wilcoxon signed-rank tests were performed. Results were considered significant when p < 0.05. For the volume assessment analyses, Lin's concordance correlation analyses (22), Bland-Altman agreement plots and paired Wilcoxon signed-rank tests were performed. Inter- and intra-reader variabilities were addressed with Bland-Altman plots.

Over the 11/82 (13%) included patients, 8 patients (age 54.9 ± 13.2 years, weight 72.8 ± 7.9 kg, 2 women) successfully completed the study (MR-HF-NIV stabilization period: 5.63 min, range [4–6]) without any adverse events. Three patients did not complete it for the following reasons: 1 could not achieve the MR-HF-NIV sequence until the end and 2 canceled the MRI session for reasons unrelated to the study.

Based on the CT gold standard analysis, 66 nodules ≥4 mm were included (mean size 6 ± 3 mm, range 4–23 mm; 22 of 4 mm, 12 of 5 mm, 11 of 6 mm and 21 > 6 mm of diameter), thus above the determined minimal sample size of 28. Three nodules were classified as subsolid and 63 as solid, among which 12 were classified as lymph nodes. While none of the subsolid nodules were identified as such in MR images, one reader identified properly 4 lymph nodes out of 12 in the VIBE-BH images and 1 lymph node out of 12 in the UTE-HF-NIV images.

UTE-HF-NIV images provided the highest detection rate for two of the three readers and for the two evaluated sizes of nodule ≥ 4 mm (n = 66) and ≥ 6 mm (n = 32) (35 and 50%, respectively), while UTE-FB provided the lowest detection rate for all readers (Table 2). False negatives were related either to the invisibility of the nodule, to a l_axis measurement <4 mm on MR images while ≥ 4 mm in CT images, or because findings were attributed to subpleural abnormalities or parenchymal bands (Figure 2). Depending on the reader, UTE-HF-NIV image analysis resulted in 4 to 9 false positives (false positive rate 6–14%) while VIBE-BH images analyses resulted in 4 to 6 false positives (false positive rate 6–9%).

With the UTE-HF-NIV, no dependencies were observed between the nodule's detection and its superior-middle-inferior location in the lung (p = 0.45, 0.55 and 1 for the three readers). One reader confirmed these results for the VIBE-BH, conversely to the two other readers who reported significant differences between superior, middle or inferior location of the nodules. Except for two readers with the UTE-HF-NIV (p = 0.002, p = 0.04), no dependencies were found between the nodule's detection and its d_pleura (p = 0.8, 0.5 and 0.8 for VIBE-BH, p = 0.05 for UTE-HF-NIV). For all readers and the two investigated techniques, there was a significant dependency between the MR nodule's detection and the nodules' measured l_axis, d_mean and V_soft (p < 0.02, Figure 3).

Given that UTE-HF-NIV and VIBE-BH had the highest detection rate among all 3 readers, the nodule size assessment was carried out with these two MR techniques. The 2D-l_axis measurements comparison resulted in a bias of 1.6 mm (p = 0.002, n = 22 nodules) between CT and VIBE-BH and 1.7 mm (p = 0.003, n = 20 nodules) between CT and UTE-HF-NIV. MR nodule volume assessment was performed on 31 nodules (mean V_soft = 94 mm³, range 12–1,068 mm³) in UTE-HF-NIV and VIBE-BH images. Regarding CT, no significant differences were observed when the same reader measured the nodule volumes in either soft or lung kernel images (p = 0.2) nor when measurements were compared between the two readers in soft kernel images (p = 0.3). When comparing nodule volume measurements between CT soft kernel and UTE-HF-NIV, Lin's concordance ρ_c and Pearson's correlation ρ values were fairly similar, suggesting no systematic bias with an excellent concordance for measured nodule volume (ρ_c = 0.980 [95%CI 0.969–0.991], ρ = 0.988, C_b = 0.992), corresponding to an overestimation of 13.2% of nodule volumes by UTE-HF-NIV compared to CT soft kernel (Figures 4A,B). On the other hand, VIBE-BH overestimated the nodule volume by 28.8%, with a significantly lower concordance (ρ_c = 0.940 [95%CI 0.913–0.968], ρ = 0.988, C_b = 0.951) as compared to UTE-HF-NIV (p < 0.05). Bland-Altman plots demonstrated almost no bias in nodule volume measurements between CT and UTE-HF-NIV as compared to VIBE-BH (bias = −0.16 mm³ and −41 mm³, respectively, Figures 4C,D).