The Siemens Biograph mMR PET/MRI system, utilized in this study, integrates simultaneous PET and MRI subsystems with 50 cm field of view (FOV) operating in 2D and 3D modes. Detector assembly for the PET sub-system part consist of 56 detector cassettes, each housing eight axially distributed 8 × 8 lutetium oxyorthosilicate crystal arrays linked to a 3 × 3 APD array for scintillation light readout. Complete camera description and geometry details are available in [19].

Patient data

This study utilized neuro PET/MR datasets from eleven patients (median [IQR] age: 70 [65, 72.5] years old, 7 Females) acquired at the Washington University Knight Alzheimer Disease Research Center (ADRC), with approval from the institutional review board and patient consent. Tri-modality brain PET/MRI/CT images were acquired using Biograph mMR PET/MRI and Biograph True Point 40 PET/CT systems (Siemens Healthcare). Head CT images were obtained with the clinical Biograph True Point 40 PET/CT system. The quantitative accuracy of the brain PET images was evaluated using four MRAC approaches and compared to CTAC as a reference. The two-point DIXON MRI sequence (DIXON) [5] segmenting head tissue as fat and water like only, the same two-point DIXON but including a skull model (DIXONbone) [20], the ultra-short echo-time MRI sequence (UTE) [21] that extract bone information from short relaxation time of protons in bone, and a DIXON-trained deep-learning-network-generated pseudo-CT map [22]. Three MRAC approaches, DIXON, DIXONbone, and UTE, are available on the mMR PET/MRI system. T1 Magnetization-prepared Rapid Acquisition Gradient Echo (MPRAGE) MRI images were processed with FreeSurfer to provide a patient-specific brain atlas of 16 regions, which were used to define lesion shape and location in the synthetic lesion insertion tool.

PET imaging

Patients were injected with an F-18-based amyloid-binding radio-ligand (Florbetapir). Data were acquired 50 min post-injection for 20 min from 10 patients, and immediately after injection for 70 min from 1 patient, using the Biograph mMR PET/MRI system. List-mode files were acquired and re-binned to sinograms using the Siemens research reconstruction software e7tools (Siemens healthcare). PET images were reconstructed using a 3D OSEM algorithm at 3 iterations, 24 subsets, and a 4 mm post-reconstruction Gaussian smoothing kernel [23]. The PET reconstructed image sizes are 344 × 344 × 127 voxels at 2.08 × 2.08 × 2.03 mm3 each.

CT imaging

Low-dose brain CT images were acquired using the CT subsystem of the Biograph TruePoint 40 PET/CT scanner at 120 kVp, 25 mAs exposure. Images were reconstructed using the filtered back-projection algorithm with H19f. The dimensions of the brain CT images are 512 × 512 × 70 voxels at 0.59 × 0.59 × 2 mm3 per voxel.

MRI imaging

Three brain MRI images were acquired using the Biograph mMR PET/MRI system using vendor-provided sequences, the standard two-point Dixon–volumetric interpolated breath-hold examination (VIBE), the high-resolution two-point Dixon CAPI, UTE, and the MPRAGE.

MRI T1-weighted brain images were acquired using a 3D MPRAGE sequence with the following imaging parameters: TE/TR = 2.95/2300 ms, TI = 900 ms, number of partitions = 176, matrix size = 240 × 256 × 176, voxel size = 1.05 × 1.05 × 1.2 mm3, acquisition time = 5 min 11 s. The T1-weighted image was used as an input to FreeSurfer to generate the patient specific brain atlas.

Attenuation mapsDIXON

The DIXON attenuation map was acquired using a vendor-provided two-point Dixon VIBE MRI sequence with a 10° flip angle (FA). At repetition time (TR), 3.6 ms, there are two echo-time TE readouts, in-phase, 2.46 ms, and out-phase, 1.23 ms, from which fat and water dominant images are generated. The acquisition time was 19 s. Four classes of tissues are generated for whole-body PET/MRI applications: air, fat, and soft tissues, to which fixed 511 keV photons attenuation coefficient were assigned. In this study, the whole brain and the skull are considered uniform soft tissues (water). The dimensions of the DIXON images are 192 × 126 × 128 voxels, and the voxel size is 2.6 × 2.6 × 3.12 mm3.

DIXONbone

The DIXONbone images were generated based on the high-resolution DIXON Controlled Aliasing in Parallel Imaging Results in Higher Acceleration (CAIPIRINHA) images (TE1/TE2/TR = 1.28/2.51/4.14 ms, FA = 10°, dimensions = 384 × 204 × 132 voxels, and voxel size = 1.30 × 1.30 × 2.02 mm, acquisition time = 39 s). Skull bones linear attenuation coefficient (LAC) replaced the soft tissues LAC in the high-resolution DIXON images. The skull bones were generated using a vendor-provided model-based bone prototype segmentation algorithm (Siemens AG, Erlangen, Germany). The first step is to generate a bone model from pre-aligned MRI images and bone masks containing continuous bone LAC at 511 keV photons. In the second step, the patient DIXON image is registered to the generated MRI model, then bone masks are registered to the patient DIXON image, segmented in the bone tissues, and brought back to the original DIXOM image space using the same transformations [20, 24].

UTE

The UTE images were generated using a vendor-provided MRI imaging sequence with a 10-degree FA, 4.64 ms TR, and 0.07 ms and 2.46 ms TE, which results in simultaneous generation of cranial bones and the brain tissues. The acquisition time was 144 s. The resulted images were segmented into two compartments, soft tissues for the whole brain and bones-tissues for the cranial bones. The size of the raw UTE images consists of 192 × 192 × 192 voxels, 1.56 × 1.56 × 1.56 mm3 per voxel.

DL-DIXON

Synthetic pseudoCT attenuation maps were generated using a deep-learning technique. A network that combines the 3D residual and UNet architectures (ResUNet) was used. Pseudo-CT images were generated from the standard in- and opposite-phase DIXON images. More details about the DL-DIXON attenuation maps generation methodology, network architecture, training, and testing datasets were published in previous work in [22].

Attenuation maps preprocessing

The four MRI-derived attenuation maps and CT images obtained directly from the Biograph mMR, and True Point 40 PET/CT systems, respectively, were resampled using nearest neighbor interpolation onto the default attenuation map gridded on 344 × 344 × 127 voxels at 2.086 × 2.086 × 2.031 mm3/voxel. The DIXON, DIXONbone, and DL-DIXON MRAC were registered to the UTE attenuation map using a 12-parameter affine registration with the FMRIB Linear Image Registration Tool (FLIRT) in the FSL toolbox [25].

The CT and pseudo-CT Hounsfield (HU) unit were converted to 511 keV linear attenuation coefficients by piecewise linear scaling [26]. The CT attenuation maps were aligned to DIXON, DIXONbone, UTE, and DL-DIXON attenuation maps using a 12-parameter affine registration with FLIRT. All MR and CT attenuation maps were then wrapped to default attenuation map space using vendor-provided e7tools software.

Pipeline description

Our pipeline, visualized in Fig. 1, combines a validated synthetic lesion insertion tool for Siemens mMR [18] and FreeSurfer framework for brain segmentation, utilizing T1-weighted MRI images to produce a patient-specific brain atlas. Brain regions of interest (ROIs), with definable activity or Standardized Uptake Value (SUV), are input into the tool, optionally using original PET image data for lesion-to-background ratio (LBR) application and smoothing via the scanner's point spread function (PSF). Subsequent to this, lesion ROI and attenuation map images are forward-projected, with the lesion ROI sinogram undergoing multiple processing stages—including voxel-wise division by extended normalization and integral factors matrices, calibration of lesions activity, calculation and addition of scatter using e7tools, and Poisson noise addition—before being added to or replacing sinogram counts in the patient PET projection space. Final image reconstruction is achieved using standard 3D-OSEM with three iterations and 21 subsets [23]. More technical details about the lesion insertion tool are presented in [18].

Fig. 1

Pipeline for the evaluation of different PET attenuation correction approaches using the synthetic lesion ROIs insertion tool and FreeSurfer

Brain regions of interest

In a typical comparison of PET/MRI to PET/CT for attenuation correction evaluation, PET emission data were reconstructed with two different attenuation maps: a specific MRAC and a CTAC. For regional brain uptake analysis, brain ROIs are delineated using manual or automatic approaches to calculate the uptake deviation from MRAC PET reconstruction relative to CTAC PET reconstruction. In the case of automatic ROIs generation, for instance, a brain atlas generated from a FreeSurfer T1 weighted MRI images with 256 × 256 × 256 voxels at ~ 1 mm3/voxel needs to be in the same space as the final reconstructed PET images. The FreeSurfer brain atlases were aligned to the PET using rigid registration with FSL's FLIRT. Figure 2a presents a 2D slice of a brain atlas superimposed on its corresponding 2D brain PET image. Brain ROIs are defined, projected to sinogram space, and reconstructed with and without considering the patient's sinogram, using a lesion insertion tool. The MRAC to CTAC bias is then compared in the inserted 16 brain ROIs and the original PET images.

Fig. 2

Examples of FreeSurfer brain atlas (a) and two spherical lesions (b and c). The atlas and the inserted lesions are superimposed over the PET image

Pipeline evaluation

Two spherical 4 mm radius lesions were inserted, one in the superior frontal cortex and another in the fusiform gyrus. Figure 2b, c show examples of the two spherical lesions inserted in the brain. The MRAC to CTAC bias is most sensitive to lesion location; a lesion inserted near the skull shows a higher bias than one inserted farther from the brain skull. PET images with inserted lesions were reconstructed with and without the brain PET background (or projections) using the four MRAC maps, DIXON, DIXONbone, UTE, DL-DIXON, and the CTAC map. In addition, PET images without the inserted lesions were also reconstructed using the same MRAC maps. PET reconstruction bias in the inserted lesions ROIs were compared across methods.

Data analysis

MRAC to CTAC bias was calculated in the inserted spherical lesions ROIs and FreeSurfer brain atlas ROIs using relative and relative absolute errors in Eq. 1 and Eq. 2, respectively. Box plots of the relative and absolute MRAC to CTAC bias were displayed for the entire patient cohort.

$$Relative \;bias = \left( - PET_ } \right)/PET_$$

(1)

$$Aboslute\; bias = \left| - PET_ } \right)} \right|/PET_$$

(2)

Statistical analyses

Statistical analyses were performed using R 4.2.0 (Foundation for Statistical Computing, Vienna, Austria). Comparisons between PET absolute bias using DL-DIXON AC maps and using other three MRAC maps were performed using paired t tests with the Benjamini–Hochberg to control for the false discovery rate in multiple comparisons.