This prospective study was approved by the institutional review board and the local ethics committee. Written informed consent was obtained from all participants prior to study inclusion. The study employs imaging data from a previously published cohort [2]. The prior report compared bone assessment of the sacroiliac (SI) joint by CT and ZTE MRI. The current study expands this by generating DL-based pCT images from the acquired ZTE MR sequences.

Individuals aged > 18 years who were referred for clinically indicated MR scans of the abdomen or pelvis between May 2019 and January 2021 were recruited. All patients considered for enrollment had undergone a CT scan covering the SI joints within 12 months of their MR examination. If patients agreed to participate in the study, an additional ZTE sequence was added to the respective standard MR protocol.

Exclusion criteria were refusal to participate in the study, pregnancy, contra-indications to MRI, any form of incomplete datasets, major artifacts due to motion or foreign bodies, or severe B1 field inhomogeneity.

All MR scans were obtained with a 3.0-T scanner (Discovery 750W plus GEM; GE Healthcare) using an abdominal coil. Different protocols were used according to the respective clinical indications. At the end of each protocol, the same ZTE sequence was acquired in each participant (oZTEo, GE Healthcare; TR, 5.1 ms; TE, ≈ 0 ms; acquisition matrix, 212 × 212 × 250; slice thickness, 1.5 mm; field of view, 320 mm; bandwidth, ± 62.5 kHz; flip angle, 2°; scan time, 4:06 min). ZTE images were acquired in an axial plane in isotropic resolution. The typical through-plane coverage was 250 slices, ranging from the 12th thoracic vertebra to the lesser trochanter.

pCT images were generated from ZTE MRI using the method previously described in [12]. The solution consists of a 2D multi-layer convolution neural network adapted to a multi-task UNet framework. The network is designed to maintain the overall structural accuracy of the image while focusing on achieving precise bone representation by learning correlated tasks: (a) image translation as the primary task, (b) bone segmentation, and (c) bone density value estimations as auxiliary tasks. Each task is optimized individually using a dedicated loss function that is customized to minimize a specific error, and the combined loss value contributes towards the overall training of the network. By separating the tasks of classification and regression, and by optimizing the network to reduce both errors simultaneously, implicit reinforcement can be achieved towards each of the correlated tasks [13]. Although the tasks are correlated, the network is expected to learn them differently from one another.

CT examinations were performed on different scanners. The majority of scans consisted of an abdominal or pelvic CT obtained with the latest generation energy-integrating dual-source scanner (Somatom Definition FLASH, Siemens Healthineers; tube voltage, 90 kVp, 100 kVp, 110 kVp, and 120 kVp; tube current, 100–150 mAs with active tube modulation); field of view, 500 mm with a matrix of 512 × 512; bone kernel (mostly Br59) for axial image reconstruction). Image data were then reformatted in all three planes with a slice thickness and increment of 1.5 mm, each.

The assessment of CT and pCT images was performed independently by two fellows in musculoskeletal radiology (J.M.G. and S.M., both 5 years of experience) after careful instructions by a senior musculoskeletal radiologist with more than 15 years of experience (R.G.). The readout of the two datasets (CT and pCT, respectively) was conducted in two different sessions, separated by 4 weeks and in different random order to avoid recall bias. Both readers were blinded to patient identification and clinical data, as well as to the results of the other datasets. CT images were evaluated using the institution’s picture archiving and communication system (DeepUnity Diagnost, version R20 XX; Dedalus S.p.A.). pCT images were viewed using Synedra View 21 (version 21.0.0.8 (× 64 edition); Synedra Information Technologies GmbH).

CT and pCT images were rated qualitatively using a 4-point Likert scale (0–3; 0 = poor, 1 = slight, 2 = good, and 3 = perfect) [14] with regard to the following parameters: sharpness of bone contour, differentiation of cortical and trabecular bone, delineation of hip joint space, delineation of SI joint space, and preservation of soft tissue boundaries. For pCT, body masking and severity of false bone classification around the pelvis (0–3; 0 = none, 1 = slight, 2 = marked, and 3 = severe) were assessed additionally. Subjective assessment confidence was rated separately for all image series (0–3; 0 = poor confidence that makes it almost impossible for assessment, 1 = low confidence that may affect the assessment, 2 = moderate confidence that does not affect the assessment negatively, and 3 = high confidence facilitating a clear assessment).

To assess the geometrical accuracy of the synthesized pCT images, the following measurements were performed in both CT and pCT images: distance between the center of the femoral heads in the axial plane, transverse (greatest width of the superior pelvic aperture) and anteroposterior (measured from the pubic symphysis to the sacral promontory) pelvic diameter, alpha angle of the right femur in the oblique axial plane, and lateral center-edge angle of the right femur in the coronal plane.

Additionally, HU values (mean, SD, minimum, and maximum) were determined using same-sized region of interest (ROI) measurements (3 mm2 for cortical bone; 15 mm2 for all other anatomic locations) in the following structures in both CT and pCT: cortical bone of the body of right ilium, trabecular bone of the body of right ilium, right gluteus maximus, subcutaneous fat adjacent to right gluteus maximus, and air close to tissue in areas that were visually free of noise.

To quantitatively assess image quality and tissue differentiation, contrast-to-noise ratios (CNR) were calculated for CT and pCT images with the following formula:

$$}\,=\,\frac}}}}_}}(A)-}}_}}(B)}}}}}}_}}}}$$

where (A) and (B) are structures in the ROI, and air is defined as pure image noise.

All statistical analyses were conducted using SPSS (version 29.0; IBM). p-values < 0.05 were considered statistically significant. Intraclass correlation coefficients (ICC) were calculated for all qualitative categories based on Likert scales to assess rating consistencies between both readers and methods. ICC estimates and their 95% confidence intervals (CI) were calculated based on a mean rating (k = 2), consistency agreement, and a two-way mixed-effects model. ICC values less than 0.50 were considered poor, between 0.50 and 0.75 moderate, between 0.75 and 0.90 good, and above 0.90 as an excellent agreement [15]. All distance and angle measurements were first evaluated regarding their normal distribution using a Shapiro–Wilk test [16]. If a normal distribution was present, a paired sample t-test was applied to evaluate differences between readers, respectively methods. If measurements did not show a normal distribution, a Wilcoxon signed rank test was calculated and on all significant results, a post-hoc Holm–Bonferroni test for multiple comparisons was performed [17].