Open-access template and database approaches for pseudo-CT generation in brain PET/MRI attenuation correction

35 magnetic resonance-based attenuation correction 36 pseudo-CT 37 open-access database 38 spatial normalization 39 image-based attenuation map generation 40 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 3

41 Positron emission tomography (PET) provides high -sensitivity molecular information in 42 neuroimaging. However, its quantitative accuracy critically depends on attenuation correction 43 (AC). Unlike PET/CT, hybrid PET/MRI systems cannot directly measure photon attenuation , 44 necessitating alternative AC strategies. While several MRI -based AC tools exist, they often 45 raise concerns regarding data privacy, long-term accessibility and reproducibility. We present 46 an open-access framework for PET AC using pseudo-CT (pCT) methods. A database of 60 47 paired MRI-CT scans was curat ed and normalized to stereotactic MNI space. T hree 48 established pCT approaches (Boston, MaxProb, UCL) were adapted to enable their use with 49 the normalized database. Method performance was assessed by comparison with a CT-based 50 AC reference in an independent validation sample of 28 subjects, using the Jaccard index and 51 [11C]DASB PET-derived serotonin transporter quantification as evaluation metrics. Boston and 52 MaxProb demonstrated consistent performance across database subsets, while t he UCL 53 achieved the closest agreement with CT-derived AC and yielded the lowest quantification error 54 (mean relative error <0.5% ). This work demonstrates the feasibility of different pCT 55 approaches using a database in MNI space thereby introducing an openly available reference 56 database to support fast, reproducible and continued integration of MRI-based AC into PET 57 reconstruction pipelines. 58 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 4

59 Positron emission tomography (PET) is an essential tool for studying both normal brain function 60 and pathological alterations across diverse patient populations. The field of molecular imaging 61 with PET now spans a broad range of radioligands, outcome measures from standardized 62 uptake value (SUV) to volumes of distributions (1) and more recently , stimulation-specific 63 activation (2–4). 64 Accurate reconstruction of radioligand concentration is critical for all these approaches, and a 65 key feature there of is attenuation correction (AC) , which requires an estimate of photon 66 attenuation in the head (5). Traditionally, AC is performed using transmission scans with e.g., 67 a 68Ge source rotated around the subject, or more recently X-ray computed tomography (CT) 68 (6,7). However, in hybrid PET/MRI systems these reference standards are not available 69 without additional scans, and standard MRI sequences cannot fully characterize tissue density. 70 To address this limitation, alternative methods have been developed to create so-called 71 pseudo-CTs (pCTs) for AC in PET/MR imaging . These fall into three broad categories: 72 segmentation-based, database-based and emission-based AC. Segmentation-based AC (8–73 10) identifies tissue types utilizing specialized MRI sequences. Database AC approaches (11–74 14) align pairs of MRI and CT from a database to the subject’s anatomy. Emission-based AC 75 estimates attenuation directly from PET data using, for example, maximum likelihood 76 algorithms (15) or deep learning methods (16). An in-depth comparison of various attenuation 77 correction methods exceeds the scope of this work, and readers are referred to other 78 publications (5,6,17–21). 79 Beyond accuracy, practical factors such as accessibility, ease of integration and adaptability 80 to evolving research needs are increasingly important aspects of software tools . The 81 neuroimaging community has embraced t hese principles th roughout widely used open 82 frameworks such as SPM 12 (https://www.fil.ion.ucl.ac.uk/spm/), AFNI 83 (https://afni.nimh.nih.gov/) and FSL (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) for processing of 84 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 5 imaging data, data management with Brain Imaging Data Structure (BIDS) (22), anatomical 85 atlases (23) as well as new approaches to kinetic modelling (24) and functional PET (25). 86 With respect to AC, several database pCT generation tools (e.g.: UCL 87 (http://niftyweb.cs.ucl.ac.uk/); (11)); MaxProb (13,26)) are available for online use. Notably the 88 RESOLUTE algorithm (10) is available on GitHub for local installation, but was found to be 89 sensitive to acquisition parameters and scanner software version (14). Online tools enable 90 creation of pCTs with minimal computational resources on the user’s end but raise potential 91 data protection concerns, as they require uploading unprocessed (i.e., non-defaced, native 92 space), individual T1 -weighted images. In addition, processing time can be long since the 93 entire database must be registered to each individual T1-weighted image. Requests from 94 several users at the same time might lead to longer processing queues, increasing waiting 95 time. Moreover, maintenance or long-term availability are not guaranteed and custom solutions 96 are not supported. 97 In this work, we present adaptations of the MaxProb, Boston and UCL database approaches 98 for pCT creation, using a database that is pre-aligned to MNI space. This strategy aims to 99 improve computational efficiency, avoids sharing raw subject data, and facilitates seamless 100 integration into existing PET/MRI reconstruction pipelines. We further examine how database 101 size influences pCT performance, which is evaluated with an independent validation sample 102 of subjects acquired within the same study but n ot included in the database . To support 103 accessibility, we provide an openly available database of MRI -CT pairs, corresponding 104 templates and a lightweight MATLAB implementation for use with the Siemens reconstruction 105 software. 106 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 6

107 Subjects and study design 108 In total, 88 subjects (mean age ± SD = 27.5 ± 8.7, 52 female) previously included in a clinical 109 trial ( EudraCT Number : 2014 -001550-42) were selected for creating pCTs (n = 60) and 110 validation of pCTs for attenuation correction (n = 28). More detailed information on the subjects 111 used for the various templates and validation can be found in Table 1. Part of the sample was 112 already published previously in a different context (27–29). 113 Subjects underwent general medic al examination at the screening visit, including routine 114 laboratory tests, electrocardiography, and neurological assessment. Exclusion criteria 115 included a history of neurological disorders, substance abuse, smoking, pregnancy, or current 116 breastfeeding. Alongside healthy volunteers (n = 51), subjects diagnosed with major 117 depressive disorder (MDD) (n = 37) were included in the analysis. The study (clinical trial 118 including attenuation correction methodology) was approved by the Ethics Committee of the 119 Medical University of Vienna (1307/2014) and procedures were carried out in accordance with 120 the Declaration 1964 of Helsinki. A ll subjects gave written informed consent. Sharing of the 121 MR-CT image pairs was approved by departments of legal affairs and data clearin g of the 122 Medical University of Vienna (2025-042). 123 Neuroimaging 124 All subjects underwent a low-dose CT scan using a Biograph TruePoint PET/CT (Siemens 125 Medical, Erlangen, Germany) (tube potential: 120 kVp, tube current: 58 mA, voxel size: 0.59 x 126 0.59 x 1.5 mm 3) as well as a PET/MR measurement with the radioligand [ 11C]DASB on a 127 Biograph mMR (Siemens Medical, Erlangen, Germany). A total calculated dose of 20 MBq/kg 128 was delivered via a bolus and constant infusion protocol (1 min bolus, 179 min infusion). The 129 125-minute PET scan started 30-45 minutes after the initial bolus. During the PET examination, 130 a T1-weighted magnetization-prepared rapid gradient-echo sequence (TE/TR = 4.21/3000 ms, 131 voxel size 1 x 1 x 1.1 mm 3) and the vendor-provided DIXON-VIBE sequence (TE1/TE2/TR = 132 1.23/2.46/3.60 ms, flip angle = 10, voxel size 2.6 x 2.6 x 3.1 mm 3) were acquired. For the 133 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 7 quantification of serotonin transporter binding, arterial blood samples were taken at minutes 134 120, 130, 140 post radioligand application. 135 Attenuation correction approaches 136 A schematic overview for the creation of attenuation maps is shown in Figure 1. 137 To implement the various pCT approaches, we curated an in-house database of MRI-CT pairs. 138 For each subject in the database , the CT scan was co-registered to th e corresponding T1 -139 weighted MR image via a rigid transformation using SPM12. Images were resliced (4th degree 140 B-Spline) and p roper alignment was visually verified. T1-weighted images were then 141 normalized to MNI space (SPM12, default settings) , using the ti ssue probability map (TPM) 142 provided in SPM12. The head holder was masked using the TPM and background was set to 143 -1024 HU. The resulting transformation was then applied to the co -registered CT. pCT 144 generation followed the algorithms described in prior work, except for having CTs in normalised 145 MNI space instead of individual space for the implementation of the modified MaxProb 146 (mMaxProb) approach. For MaxProb (13,26), each CT was segmented into three tissue 147 classes by intensity (I) thresholding: I < -500 HU (air), -500 HU ≤ I < 300 HU (soft tissue) and 148 I ≥ 300 HU (bone) (13,30). For each voxel, the most prevalent tissue class at that location 149 across all database CTs was determined. The pCT template was then generated by averaging 150 Hounsfield units over voxels belonging to the most prevalent tissue class. The Boston method 151 (12) differs from MaxPro b, as implemented here , only in the averaging step: all voxels 152 contribute to the voxel-wise mean, regardless of tissue class . For both methods, attenuation 153 maps were scaled to linear attenuation coefficients (LAC) at 511 keV using bilinear scaling 154 (31). For creation of Boston and mMaxProb pCTs, each subject’s T1 -weighted image was 155 normalized to MNI space, and the inverse transform was then applied to the template for 156 transformation to individual subject space. Database CTs were first normalized to MNI space, 157 ensuring that the only subject -specific deformation applied to the resulting attenuation maps 158 was the inverse transformation back to native space. To improve computational efficiency, the 159 MNI-space CTs were subsequently aggregated into template volumes. Accordingly, both the 160 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 8 Boston method and the mMaxProb approach are classified as template -based in this 161 manuscript. 162 To investigate the effect of database size on performance, we created mMaxProb and Boston 163 templates from subsets of 5, 10, 20, 40 and 60 subjects. Throughout the manuscript, the 164 database size will be referred to as an index (e.g. mMaxProbdb20). To minimize potential 165 confounding effects, each subset was balanced for sex distribution and age range. In absence 166 of an index referring to the specific database size, the modified implementation of MaxProb 167 will be referred to as “mMaxProb” to distinguish it from the original implementation. 168 Additionally, the UCL approach (11) for pCT generat ion was adapted to work with the MNI 169 space database. A convolution-based fast local normalized correlation coefficient (LNCC) (32) 170 was calculated between a subject ’s and each database T1-weighted image. Subsequently, 171 LNCCs were used for r anking each voxel in the database images . Weights (W) were then 172 calculated based on rankings (R) according to Eq1. Applying these weights to the database 173 CTs yielded pCTs in MNI space. Pseudo-CTs were transformed to individual space in the same 174 manner as for the template-based pCTs. This implementation of the UCL approach uses the 175 full database size of 60 subjects. Unlike the MaxProb and Boston implementation, database 176 CTs cannot be combined into a template, because the weighted mean of database CTs is 177 subject-specific. 178 𝑊 = 𝑒−0.5𝑅 Eq1 Since the field of view (FOV) of an image normalized in SPM12 does not cover the entire head 179 and neck, we supplemented these pCTs with the vendor provided attenuation map (DIXON) 180 (8), as described in (6). We intentionally did not replace the corresponding region in the CT, in 181 order to capture a possible effect and to recreate a realistic use case. Finally, all pCTs were 182 resliced (4th degree B-Spline) to the DIXON attenuation map and converted to DICOM format 183 for integration into the Siemens reconstruction pipeline. To preserve the metadata contained 184 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 9 in the original DIXON attenuation map DICOM file, only image data was modified by using the 185 dcmodify tool from the DICOM toolkit (DCMTK) package. 186 The three pCT approaches were validated in an independent samp le, whose subjects were 187 not part of the database. As a reference standard, attenuation maps were generated from each 188 subject’s CT image (10,31). Analogous to the pCTs, CT scans were scaled to LACs at 5 11 189 keV, the headrest was masked, CTs were coregistered to subject’s T1-weighted image and 190 resliced to match the dimensions of the DIXON attenuation map. 191 Image Reconstruction 192 PET images were reconstructed using the Siemens reconstruction software (e7tools, Siemens 193 Medical Solutions, Knoxville, USA) as described previously (33). Three consecutive 10-minute 194 frames were reconstructed (ordinary Poisson -ordered subset expectation maximization, 3 195 iterations, 21 subsets, post-reconstruction filter 5 mm Hann, matrix size 344 x 344 x 127, voxel 196 size 2.09 x 2.09 2.03 mm 3, Zoom 1) in tracer equilibrium, starting 95 minutes after PET start. 197 Attenuation correction was performed with each of the attenuation maps : CT, UCL, 198 mMaxProbdb5-db60 Boston db5-db60). 199 Data processing and Quantification 200 PET data were processed using SPM12 with default parameters , unless otherwise specified. 201 The three reconstructed PET frames were realigned (quality = 1, register to mean) to correct 202 for head motion and then co-registered to the T1-weighted image. The normalization to MNI 203 space was estimated with the T1-weighted image and then applied to the PET data. Imaging 204 data was processed using SPM12 and MATLAB R2018b (The Mathworks Inc., Natick, MA, 205 USA). 206 pCT performance was evaluated by quantifying serotonin transporter (SERT) binding from 207 PET data, reconstructed with each attenuation map. Total volume of distribution (VT = CT/CP) 208 (1) was quantified using an equilibrium method (33–35). Accordingly, mean activity 209 concentrations in radioligand equilibrium from the respective compartments were obtained for 210 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 10 tissue (C T) and plasma (C P). Regional values were extrac ted for thalamus, hippocampus, 211 amygdala, frontal lobe, temporal lobe, parietal lobe and occipital lobe , defined using the 212 Harvard-Oxford atlas as provided with FSL (36), and an in-house atlas was used for cerebellar 213 grey matter (37) 214 Statistical Analysis 215 Performance of different attenuation correction approaches was analyzed in two steps. First, 216 similarity between each pCT and the reference CT attenuation map was assessed using the 217 Jaccard index (38). Separate indices were calculated for bone (LAC > 0.1cm-1) and soft tissue 218 (0.05 cm-1 < LAC ≤ 0.1 cm-1) (7). Analyses were performed for both the full FOV as well as only 219 the area covered by the original pCTs. Second, the effect of attenuation correction approaches 220 on VT was evaluated by mean relative error ( 𝑉T(CT)−VT(𝑝𝐶𝑇) VT(𝐶𝑇) ) and mean absolute error relative 221 to the CT-based attenuation correction. A repeated measures analysis of variance (rmANOVA) 222 was performed for each ROI, to compare relative errors in VT between the three pCT 223 approaches. For Boston and mMaxProb, only the best performing database size was included. 224 Bonferroni correction was used to correct for repeated testing (9 ROIs). Paired sample t-tests 225 were used for post-hoc comparison in ROIs where the rmANOVA results were significant. 226 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 11

227 A database of 60 MRI-CT pairs was curated and transformed to MNI space . Templates for 228 mMaxProb and Boston pCTs were created for five database sizes (n = 60, 40, 20, 10, 5). The 229 subcohorts did not differ significantly with regard to mean age (one-way ANOVA: p = 0.89, F 230 = 0.275) or sex (chi-squared test: p = 0.98, χ²(4) = 0.39). 231 A visual comparison between CT and pCTs for a representative subject is shown in Figure 2. 232 The template-based pCT Methods (Boston, mMaxProb) exhibited greater blurring a t larger 233 database sizes. In general, the greatest deviations from the reference CT were present in the 234 posterior part of the skull and the paranasal sinuses. In these regions, mMaxProb pCTs tended 235 to underestimate LAC values, ignoring the fine bone structu res. Similarly, Boston and UCL 236 failed to capture this complex combination of air, tissue and bone. UCL best captured variations 237 within the skull. 238 As assessed by the Jaccard index, similarity between pCTs in the segmentation of bone and 239 soft tissue was comparable for the mMaxProb and Boston approach and consistent across 240 database sizes, while the UCL pCTs perform ed numerically better, particularly for the bone 241 segment (rmANOVA followed by pairwise comparison, all p < 0.001) . As expected , the 242 observed correspondence between pCT and CT was reduced when including the neck region 243 that was supplemented by the DIXON attenuation map in the pCTs. A comprehensive overview 244 of these results can be found in Table 2. 245 Mean relative error in SERT VT (Table 3) was calculated between PET reconstructions using 246 CT-based and pCT-based attenuation correction. Quantification errors were generally low and 247 distributed around zero with a small negative bias in most ROIs, indicating an overestimation 248 of VT reconstructed with pCTs. Among the pCT approaches, UCL performed best, followed by 249 mMaxProbdb5 and Boston db20, all showing small relative mean error below 0.5% across the 250 brain. No clear trend of database size and pCT performance was evident in the relative errors. 251 Boxplots of mean relative quantification errors are provided in Figure 3 for the best performing 252 pseudo-CTs (UCL, mMaxProbdb5, Bostondb20). Boxplots for all database sizes as well as mean 253 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 12 absolute errors are provided in the supplement. The rmANOVA yielded significant results in 254 amygdala, thalamus, cerebellar grey matter and parietal lobe. The post-hoc analysis showed 255 significantly better performance of Bostondb20 pCTs compared to mMaxProbdb5 in amygdala (p 256 < 0.001), thalamus (p < 0.001) and CGM (p < 0.001) and better performance in CGM than UCL 257 (p = 0.02) . mMaxProbdb5 pCTs resulted in lower quantification errors in the parietal lobe 258 compared to UCL (p = 0.0 4) and Bostondb20 (p < 0.001). UCL performed better than 259 mMaxProbdb5 in CGM (p < 0.001). All tests were corrected for multiple testing using Bonferroni 260 correction and were evaluated at a significance level of 0.05. 261 Processing time was reduced compared to the original algorithms by having the database and 262 templates in MNI space. This reduced the number of spatial transformations per pCT to one, 263 whereas in the original individual -space implementation of MaxProb and UCL , the number 264 scales linearly with database size. When using the full database size, processing time is 265 decreased by a factor of twenty. All three adapted pCT appr oaches perform similarly with 266 regard to computational speed. 267 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 13

268 This work adapts the MaxProb, Boston and UCL algorithms for pCT generation in MNI space 269 and provides open access to the associated templates, database and code, to facilitate local 270 implementation, ensure availability , and resolve legal barriers . The proposed approach es 271 achieve comparable performance in the quantification of SERT VT while substantially reducing 272 computational demands compared to the original implementations. 273 In terms of segmentation overlap, UCL achieved the highest Jaccard index values for both 274 bone and soft tissue, consistent with visual inspection of the attenuation maps. Since our 275 SPM12-based pCTs do not cover the neck region, similarity to the reference CT was evaluated 276 both for the entire head , supplemented with the DIXON attenuation map , and for the brain 277 region alone. As expected, inclusion of the neck region reduced similarity scores , since the 278 DIXON attenuation map lacks a bone component (8). The l argest deviations in VT were 279 observed along the skull, where inaccuracies in LAC estimation can strongly im pact 280 quantification. The paranasal sinuses also posed a challenge, with prior reports noting 281 variability in sinus size and composition (air vs. fluid) (26) and proposing region-specific masks 282 to better capture the air–bone–soft tissue interface (10). 283 In general all three pCTs approaches yield low qu antification errors (< 0.5% whole brain 284 average for UCL, mMaxProbdb5, Bostondb20). Still, differences between pCTs were observed. 285 These were small in magnitude but consistent across subjects as indicated by the rmANOVA. 286 One of the sources for the observed differences stems from the averaging of the database in 287 both template-based approaches. Specifically, Boston smoothed the entire image, whereas 288 MaxProb only smoothed within tissue classes (bone, soft tissue or air), preserving sharp tissue 289 boundaries. These differences arise from the underlying averaging strategies: Boston 290 averages all database CT voxels, while MaxProb averages only within the most prevalent 291 tissue class. Neither mMaxProb nor Boston capture inter-individual variability in bone contour 292 variability as well as UCL. This is partly explained by the number of subjects contributing to 293 each voxel: for MaxProb, at least one third of the database contributes, whereas for Boston, 294 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 14 all subjects contribute equally . UCL pCTs are less blurred due to the expo nential weighting, 295 and larger databases should further improve performance as more individual variation is 296 captured. In contrast larger databases increase blur ring in Boston and MaxProb pCTs. It is 297 noteworthy that the re -implementation of the MaxProb metho d within MNI space introduces 298 methodological deviations from the original formulation. The initial approach entails direct 299 registration of database MRI-CT pairs to native subject space, with subsequent label fusion 300 performed in that same space , thus yielding subject -specific pCTs. The use of multiple 301 independent registrations in the multi -atlas-based original approach enhances robustness 302 (39). 303 In comparison to the original reports (11,12,26), our results are broadly consistent with 304 previously observed performance trends for MaxProb, Boston, and UCL methods. Our results 305 show UCL achieving the highest Jaccard indices and lowest VT errors. Moreover, (13) also 306 demonstrated that the improvement in performance by increasing database size quickly levels 307 out, which is consistent with our results. Quantitative accuracy in our study, low bias and small 308 relative error compared with CT -based AC, are also comparable to that reported in a large 309 multi-centre evaluation (6), supporting the generalizability of these methods when adapted to 310 MNI space. Notably, a spatially normalized database has several advantages and aligns with 311 the FAIR principles. First , it achieved comparable performance at substantially reduced 312 computational cost . Second, the database resolves data protection limitations inherent to 313 online-only tools such as the original UCL implementation operating with T1-weighted images 314 in native space where original faces may be reconstructed (40,41). That is, the non -linear 315 spatial transformations applied during database generation make re-identification of subjects 316 unlikely as transformation parameters are not publicly available. Third, this enables making the 317 entire database openly accessible, which in turn promotes open software solutions and 318 guarantees availability as the database may be stored locally. Fourth, independence from the 319 provider enhances reproducibility of data analyses. 320 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 15

321 Our implementation was validated using [11C]DASB dynamic PET data. While we acknowledge 322 the diversity of PET radioligands, it is important to note that these methods are adaptations of 323 established methods and the original publications show similar results and have since been 324 replicated (6), emphasizing robustness across diverse datasets. While structural brain 325 alterations have been reported in MDD (42,43), expected variations in LACs at 511 keV (44) 326 are not anticipated to significantly impact AC. The spatial coverage of the pseudo-CT templates 327 corresponds to that of the SPM12 tissue probability map, ending below the nose and excluding 328 the neck region. This region was not replaced with the DIXON attenuation map in the CT 329

s tandard, yet performance remained acceptable. We also note that scanner -330 integrated DIXON-based pCT generation tools (45), which were unavailable for our dataset, 331 could further enhance the presented methods by supplementing the template -based 332 attenuation map with more accurate subject -specific attenuation information. Dense hair 333 (braids, dreadlocks) is not accounted for in all pCT approaches and has been shown to 334 introduce localized errors in the PET reconstru ction (46). As with all database AC methods, 335 accurate spatial normalization is a prerequisite. These approaches may perform poorly in 336 subjects with major structural abnormalities (e.g., lesions, malformations). 337

338 We present an openly available database of CT-T1 image pairs in MNI space , associated 339 templates, and a streamlined pCT generation pipeline for PET/MRI attenuation correction. This 340 approach resolves key data protec tion concerns, minimizes c omputational costs , and 341 facilitates integration into custom reconstruction workflows . The local installation mode 342 ensures long-term accessibility and adaptability to evolving research needs. 343 Beyond PET/MRI, the pseudo-CT tool introduced in this work may support other neuroimaging 344 applications, such as planning functional transcranial ultrasound stimulation (47), as well as 345 simulations for transcranial magnetic stimulation (48,49). Future developmen ts should 346 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 16 incorporate subject -specific information, such as implants or EEG electrode positions, to 347 further enhance accuracy. 348 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 17

349 We thank the graduated team members and the diploma students of the Neuroimaging Lab 350 (NIL, headed by R. Lanzenbe rger) as well as the clinical colleagues from the Department of 351 Psychiatry and Psychotherapy of the Medical University of Vienna for clinical and/or 352 administrative support. In details, we would like to thank S. Kasper, D. Rujescu for their 353 medical support , L. Rischka, G.M. James for technical support , N. Berroterán -Infante, W. 354 Wadsak, M. Mitterhauser for administrative support and C. Vraka and C. Philippe for 355 radioligand synthesis. 356 AUTHOR CONTRIBUTIONS 357 Study design R.L., A.H. Data acquisition M.M., L.R.S., G.M.G., G.G., M.B.R.

C.M., A.H., M.B.R. Data Analysis C.M. Manuscript preparation – Writing C.M., A.H., M.B.R Manuscript preparation - Reviewing C.M., M.M., I.M., L.R.S., L.N., G.M.G., G.G., M.H., N.C., A.HAM., R.L., A.H., M.B.R. STATEMENTS AND DECLARATIONS 358 Ethical considerations 359 The study was approved by the Ethics Committee of the Medical University of Vienna 360 (1307/2014) and procedures were carried out in accordance with the Declaration of Helsinki. 361 Consent to participate 362 All subjects gave written informed consent after detailed explanation of the study procedures. 363 Consent for publication 364 Not applicable. 365 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 18 Declaration of conflicting interest 366 Without any relevance to this work, R. Lanzenberger received investigator -initiated research 367 funding from Siemens Healthcare regarding clinical research using PET/MR and travel grants 368 and/or conference speaker honoraria from Bruker BioSpin, Shire, AstraZeneca, Lundbeck A/S, 369 Dr. Willmar Schwabe GmbH, Orphan Pharmaceuticals AG, Janssen -Cilag Pharma GmbH, 370 Heel and Roche Austria GmbH. in the years before 2020. He has been a shareholder of the 371 start-up company BM Health GmbH, Austria since 2019. With reference to this work, A. 372 Hammers, N. Costes and I. Mérida received funding from Siemens for I. Mérida’s PhD. A . 373 Hammers has received speaker honoraria from Siemens Healthineers. The other authors do 374 not report any conflict of interest . M. Hacker received consulting fees and/or honoraria from 375 Bayer Healthcare BMS, Eli Lilly, EZAG, GE Healthcare, Ipsen, ITM, Janssen , Roche, and 376 Siemens Healthineers. 377 Funding statement 378 This research was funded in whole or in part by the Austrian Science Fund (FWF) 379 [10.55776/KLI1006, PI R. Lanzenberger ; DOI 10.55776/KLI1151, PI A. Hahn; 380 10.55776/KLI551, PI S. Kasper]. M. Murgaš was funded by the Austrian Science Fund (FWF) 381 [Grant number DOC 33-B27, Supervisor R. Lanzenberger]. For open access purposes, the 382 author has applied a CC BY public copyright license to any author accepted manuscript 383 version arising from this submission. Christian Milz is a recipient of a DOC Fellowship (27221) 384 from the Austrian Academy of Sciences at the Department of Psychiatry and Psychotherapy, 385 Medical University of Vienna. 386 Data availability 387 Raw data will not be publicly available due to reasons of data protection. Processed data and 388 custom code can be obtained from the corresponding author with a data -sharing agreement, 389 approved by the departments of legal affairs and data clearing of the Medical University of 390 Vienna. The Matlab code for creation of pseudo -CTs is available at 391 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 19 https://github.com/NeuroimagingLabsMUV/Milz_2025_pseudoCT under a CC BY -NC-SA 392 license. Imaging data, including templates and individual CTs and T1 -weighted MRI scans in 393 MNI space, are available at https://doi.org/10.5281/zenodo.17953384 under a CC BY-NC-ND 394 license. 395 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 20

396 1. Innis RB, Cunningham VJ, Delforge J, Fujita M, Gjedde A, Gunn RN, et al. Consensus 397 nomenclature for in vivo imaging of reversibly binding radioligands. J Cereb Blood Flow Metab. 398 2007 Sept 7;27(9):1533–9. 399 2. Villien M, Wey HY, Mandeville JB, Catana C, Polimeni JR, Sander CY, et al. Dynamic 400 functional imaging of brain glucose utilization using fPET-FDG. NeuroImage. 2014 Oct 401 15;100:192–9. 402 3. Hahn A, Gryglewski G, Nics L, Hienert M, Rischka L, Vraka C, et al. Quantification of Task-403 Specific Glucose Metabolism with Constant Infusion of 18F-FDG. J Nucl Med Off Publ Soc Nucl 404 Med. 2016 Dec 1;57(12):1933–40. 405 4. Jamadar SD, Ward PGD, Li S, Sforazzini F, Baran J, Chen Z, et al. Simultaneous task-based 406 BOLD-fMRI and [18-F] FDG functional PET for measurement of neuronal metabolism in the 407 human visual cortex. NeuroImage. 2019 Apr 1;189:258–66. 408 5. Martinez-Möller A, Nekolla SG. Attenuation correction for PET/MR: Problems, novel approaches 409 and practical solutions. Z Für Med Phys. 2012 Dec 1;22(4):299–310. 410 6. Ladefoged CN, Law I, Anazodo U, St. Lawrence K, Izquierdo-Garcia D, Catana C, et al. A multi-411 centre evaluation of eleven clinically feasible brain PET/MRI attenuation correction techniques 412 using a large cohort of patients. NeuroImage. 2017 Feb 15;147:346–59. 413 7. Ladefoged CN, Andersen FL, Andersen TL, Anderberg L, Engkebølle C, Madsen K, et al. 414 DeepDixon synthetic CT for [18F]FET PET/MRI attenuation correction of post-surgery glioma 415 patients with metal implants. Front Neurosci [Internet]. 2023 Apr 6 [cited 2024 Nov 12];17. 416 Available from: 417 https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2023.1142383/full 418 8. Martinez-Möller A, Souvatzoglou M, Delso G, Bundschuh RA, Chefd’hotel C, Ziegler SI, et al. 419 Tissue Classification as a Potential Approach for Attenuation Correction in Whole-Body 420 PET/MRI: Evaluation with PET/CT Data. J Nucl Med. 2009 Apr 1;50(4):520–6. 421 9. Catana C, Kouwe A van der, Benner T, Michel CJ, Hamm M, Fenchel M, et al. Toward 422 Implementing an MRI-Based PET Attenuation-Correction Method for Neurologic Studies on the 423 MR-PET Brain Prototype. J Nucl Med. 2010 Sept 1;51(9):1431–8. 424 10. Ladefoged CN, Benoit D, Law I, Holm S, Kjær A, Højgaard L, et al. Region specific optimization 425 of continuous linear attenuation coefficients based on UTE (RESOLUTE): application to PET/MR 426 brain imaging. Phys Med Biol. 2015 Sept;60(20):8047. 427 11. Burgos N, Cardoso MJ, Thielemans K, Modat M, Pedemonte S, Dickson J, et al. Attenuation 428 Correction Synthesis for Hybrid PET-MR Scanners: Application to Brain Studies. IEEE Trans 429 Med Imaging. 2014 Dec;33(12):2332–41. 430 12. Izquierdo-Garcia D, Hansen AE, Förster S, Benoit D, Schachoff S, Fürst S, et al. An SPM8-based 431 approach for attenuation correction combining segmentation and nonrigid template formation: 432 application to simultaneous PET/MR brain imaging. J Nucl Med Off Publ Soc Nucl Med. 2014 433 Nov;55(11):1825–30. 434 13. Mérida I, Costes N, Heckemann RA, Drzezga A, Förster S, Hammers A. Evaluation of several 435 multi-atlas methods for PSEUDO-CT generation in brain MRI-PET attenuation correction. In: 436 2015 IEEE 12th International Symposium on Biomedical Imaging (ISBI) [Internet]. 2015 [cited 437 2024 Nov 26]. p. 1431–4. Available from: https://ieeexplore.ieee.org/document/7164145 438 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 21 14. Ladefoged CN, Hansen AE, Henriksen OM, Bruun FJ, Eikenes L, Øen SK, et al. AI-driven 439 attenuation correction for brain PET/MRI: Clinical evaluation of a dementia cohort and 440 importance of the training group size. NeuroImage. 2020 Nov 15;222:117221. 441 15. García-Pérez P, España S. Simultaneous emission and attenuation reconstruction in time-of-flight 442 PET using a reference object. EJNMMI Phys. 2020 Jan 13;7(1):3. 443 16. Liu F, Jang H, Kijowski R, Zhao G, Bradshaw T, McMillan AB. A deep learning approach for 444 18F-FDG PET attenuation correction. EJNMMI Phys. 2018 Nov 12;5(1):24. 445 17. Bezrukov I, Mantlik F, Schmidt H, Schölkopf B, Pichler BJ. MR-Based PET Attenuation 446 Correction for PET/MR Imaging. Semin Nucl Med. 2013 Jan 1;43(1):45–59. 447 18. Berker Y, Li Y. Attenuation correction in emission tomography using the emission data—A 448 review. Med Phys. 2016;43(2):807–32. 449 19. Teuho J, Torrado-Carvajal A, Herzog H, Anazodo U, Klén R, Iida H, et al. Magnetic Resonance-450 Based Attenuation Correction and Scatter Correction in Neurological Positron Emission 451 Tomography/Magnetic Resonance Imaging—Current Status With Emerging Applications. Front 452 Phys [Internet]. 2020 Jan 29 [cited 2024 Nov 26];7. Available from: 453 https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2019.00243/full 454 20. Chen Y, An H. Attenuation Correction of PET/MR Imaging. Magn Reson Imaging Clin N Am. 455 2017 May 1;25(2):245–55. 456 21. Krokos G, MacKewn J, Dunn J, Marsden P. A review of PET attenuation correction methods for 457 PET-MR. EJNMMI Phys. 2023 Sept 11;10:52. 458 22. Galassi A, Norgaard M, Thomas AG, Gonzalez-Escamilla G, Svarer C, Rorden C, et al. 459 PET2BIDS: a library for converting Positron Emission Tomography data to BIDS. J Open Source 460 Softw. 2024;9(100):6067. 461 23. Hammers A, Allom R, Koepp MJ, Free SL, Myers R, Lemieux L, et al. Three‐dimensional 462 maximum probability atlas of the human brain, with particular reference to the temporal lobe. 463 Hum Brain Mapp. 2003 May 30;19(4):224–47. 464 24. Matheson GJ, Ogden RT. Multivariate analysis of PET pharmacokinetic parameters improves 465 inferential efficiency. EJNMMI Phys [Internet]. 2023 Dec 1 [cited 2024 Feb 5];10(1). Available 466 from: /pmc/articles/PMC10008760/ 467 25. Hahn A, Reed MB, Milz C, Falb P, Murgaš M, Lanzenberger R. A unified approach for 468 identifying PET-based neuronal activation and molecular connectivity with the functional PET 469 toolbox. J Cereb Blood Flow Metab. 2025 Sept 8;0271678X251370831. 470 26. Mérida I, Reilhac A, Redouté J, Heckemann RA, Costes N, Hammers A. Multi-atlas attenuation 471 correction supports full quantification of static and dynamic brain PET data in PET-MR. Phys 472 Med Biol. 2017 Apr 7;62(7):2834–58. 473 27. Ponce de León M, Murgaš M, Silberbauer LR, Hacker M, Gryglewski G, Hahn A, et al. 474 Simplified methods for SERT occupancy estimation measured with [11C]DASB PET bolus plus 475 infusion. NeuroImage. 2025 May 1;311:121208. 476 28. Gryglewski G, Klöbl M, Berroterán-Infante N, Rischka L, Balber T, Vanicek T, et al. Modeling 477 the acute pharmacological response to selective serotonin reuptake inhibitors in human brain using 478 simultaneous PET/MR imaging. Eur Neuropsychopharmacol. 2019 June 1;29(6):711–9. 479 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 22 29. Silberbauer LR, Rischka L, Vraka C, Hartmann AM, Godbersen GM, Philippe C, et al. ABCB1 480 variants and sex affect serotonin transporter occupancy in the brain. Mol Psychiatry. 2022 481 Nov;27(11):4502–9. 482 30. Kinahan PE, Hasegawa BH, Beyer T. X-ray-based attenuation correction for positron emission 483 tomography/computed tomography scanners. Semin Nucl Med. 2003 July 1;33(3):166–79. 484 31. Carney JPJ, Townsend DW, Rappoport V, Bendriem B. Method for transforming CT images for 485 attenuation correction in PET/CT imaging. Med Phys. 2006;33(4):976–83. 486 32. Cachier P, Bardinet E, Dormont D, Pennec X, Ayache N. Iconic feature based nonrigid 487 registration: the PASHA algorithm. Comput Vis Image Underst. 2003 Feb 1;89(2):272–98. 488 33. Rischka L, Gryglewski G, Berroterán-Infante N, Rausch I, James GM, Klöbl M, et al. Attenuation 489 Correction Approaches for Serotonin Transporter Quantification With PET/MRI. Front Physiol. 490 2019 Nov 22;10:1422. 491 34. Ginovart N, Wilson AA, Meyer JH, Hussey D, Houle S. Positron Emission Tomography 492 Quantification of [11C]-DASB Binding to the Human Serotonin Transporter: Modeling 493 Strategies. J Cereb Blood Flow Metab. 2001 Nov 1;21(11):1342–53. 494 35. Gryglewski G, Rischka L, Philippe C, Hahn A, James GM, Klebermass E, et al. Simple and rapid 495 quantification of serotonin transporter binding using [11C]DASB bolus plus constant infusion. 496 NeuroImage. 2017 Apr 1;149:23–32. 497 36. Desikan RS, Ségonne F, Fischl B, Quinn BT, Dickerson BC, Blacker D, et al. An automated 498 labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions 499 of interest. NeuroImage. 2006 July 1;31(3):968–80. 500 37. Savli M, Bauer A, Mitterhauser M, Ding YS, Hahn A, Kroll T, et al. Normative database of the 501 serotonergic system in healthy subjects using multi-tracer PET. NeuroImage. 2012 Oct 502 15;63(1):447–59. 503 38. Jaccard P. Étude comparative de la distribution florale dans une portion des Alpes et du Jura. Bull 504 Société Vaudoise Sci Nat. 1901;37(142):547. 505 39. Heckemann RA, Hajnal JV, Aljabar P, Rueckert D, Hammers A. Automatic anatomical brain MRI 506 segmentation combining label propagation and decision fusion. NeuroImage. 2006 Oct 507 15;33(1):115–26. 508 40. Bischoff‐Grethe A, Ozyurt IB, Busa E, Quinn BT, Fennema‐Notestine C, Clark CP, et al. A 509 technique for the deidentification of structural brain MR images. Hum Brain Mapp. 2007 Feb 510 12;28(9):892–903. 511 41. Schwarz CG, Kremers WK, Arani A, Savvides M, Reid RI, Gunter JL, et al. A face-off of MRI 512 research sequences by their need for de-facing. NeuroImage. 2023 Aug 1;276:120199. 513 42. Schmaal L, Veltman DJ, van Erp TGM, Sämann PG, Frodl T, Jahanshad N, et al. Subcortical 514 brain alterations in major depressive disorder: findings from the ENIGMA Major Depressive 515 Disorder working group. Mol Psychiatry. 2016 June;21(6):806–12. 516 43. Brandl F, Weise B, Mulej Bratec S, Jassim N, Hoffmann Ayala D, Bertram T, et al. Common and 517 specific large-scale brain changes in major depressive disorder, anxiety disorders, and chronic 518 pain: a transdiagnostic multimodal meta-analysis of structural and functional MRI studies. 519 Neuropsychopharmacology. 2022 Apr;47(5):1071–80. 520 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 23 44. Hubell JH, Seltzer SM. X-Ray Mass Attenuation Coefficients. Gaithersburg: National Institute of 521 Standards and Technology [Internet]. 2009 Sept 17 [cited 2025 Oct 3]; Available from: 522 https://www.nist.gov/pml/x-ray-mass-attenuation-coefficients 523 45. Paulus DH, Quick HH, Geppert C, Fenchel M, Zhan Y, Hermosillo G, et al. Whole-Body 524 PET/MR Imaging: Quantitative Evaluation of a Novel Model-Based MR Attenuation Correction 525

Including Bone. J Nucl Med. 2015 July 1;56(7):1061–6. 526 46. Mackewn JE, Stirling J, Jeljeli S, Gould SM, Johnstone RI, Merida I, et al. Practical issues and 527

of brain attenuation correction on a simultaneous PET-MR scanner. EJNMMI Phys. 528 2020 May 5;7(1):24. 529 47. Yaakub SN, White TA, Kerfoot E, Verhagen L, Hammers A, Fouragnan EF. Pseudo-CTs from 530 T1-weighted MRI for planning of low-intensity transcranial focused ultrasound neuromodulation: 531 An open-source tool. Brain Stimul Basic Transl Clin Res Neuromodulation. 2023 Jan 1;16(1):75–532 8. 533 48. Gomez LJ, Dannhauer M, Koponen LM, Peterchev AV. Conditions for numerically accurate TMS 534 electric field simulation. Brain Stimulat. 2020 Jan 1;13(1):157–66. 535 49. Windhoff M, Opitz A, Thielscher A. Electric field calculations in brain stimulation based on finite 536 elements: An optimized processing pipeline for the generation and usage of accurate individual 537 head models. Hum Brain Mapp. 2013;34(4):923–35. 538 539 540 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 24 TABLES 541 n Sex (f/m) mean Age ± SD HC/MDD db5 5 3/2 25.2 ± 7.3 3/2 db10 10 5/5 27.6 ± 5.5 8/2 db20 20 12/8 25.8 ± 7.5 14/6 db40 40 23/17 26.0 ± 6.5 25/15 db60 60 36/24 25.4 ± 7.5 38/22 validation set 28 16/12 32.1 ± 11.6 13/15 542 Table 1: Demographic information on d atabase subsets and validation set. Subjects in the 543 validation set are not included in the database. 544 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 25 545 pCT J(BoneH) J(BoneHN) J(TissueH) J(TissueHN) UCL 0.71 ± 0.07 0.58 ± 0.05 0.80 ± 0.03 0.72 ± 0.03 MaxProbdb60 0.66 ± 0.07 0.54 ± 0.05 0.78 ± 0.03 0.72 ± 0.03 MaxProbdb40 0.66 ± 0.07 0.54 ± 0.05 0.78 ± 0.03 0.72 ± 0.03 MaxProbdb20 0.65 ± 0.06 0.53 ± 0.04 0.78 ± 0.03 0.72 ± 0.03 MaxProbdb10 0.65 ± 0.06 0.53 ± 0.04 0.78 ± 0.03 0.72 ± 0.03 MaxProbdb5 0.64 ± 0.06 0.53 ± 0.05 0.78 ± 0.03 0.72 ± 0.03 Bostondb60 0.66 ± 0.07 0.55 ± 0.05 0.77 ± 0.03 0.71 ± 0.03 Bostondb40 0.66 ± 0.07 0.55 ± 0.05 0.77 ± 0.03 0.71 ± 0.03 Bostondb20 0.66 ± 0.06 0.54 ± 0.05 0.77 ± 0.03 0.71 ± 0.03 Bostondb10 0.65 ± 0.06 0.54 ± 0.05 0.77 ± 0.03 0.71 ± 0.03 Bostondb5 0.65 ± 0.06 0.53 ± 0.05 0.78 ± 0.03 0.72 ± 0.03 546 Table 2: Jaccard index (J) between computed tomography (CT) attenuation map and 547 pseudo-CTs (pCTs). Similarity is evaluated for bone (LAC > 0.1cm-1) and soft tissue (0.05 548 cm-1 < LAC ≤ 0.1 cm-1) in the head region (H) without the neck and across the entire 549 attenuation map (HN). Results are reported as mean ± SD. 550 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 26 thalamus hippocampus amygdala cerebellar grey matter frontal lobe temporal lobe parietal lobe occipital lobe whole brain UCL -0.48±1.05 -0.47±1.19 -0.29±1.25 -1.05±2.41 -0.06±1.70 -0.42±2.23 0.36±2.38 0.30±1.93 -0.03±2.77 MaxProbdb60 -1.10±1.36 -1.21±1.49 -0.95±1.55 -1.95±2.87 -1.34±2.29 -1.60±2.68 -0.75±2.92 -0.93±2.37 -1.19±3.65 MaxProbdb40 -1.28±1.32 -1.36±1.47 -1.06±1.52 -2.23±2.90 -1.46±2.27 -1.79±2.70 -0.87±2.92 -1.20±2.38 -1.35±3.56 MaxProbdb20 -0.74±1.33 -0.95±1.43 -0.58±1.54 -0.94±2.84 -0.99±2.26 -1.30±2.69 0.18±2.94 0.14±2.35 -0.60±3.64 MaxProbdb10 -1.05±1.36 -1.07±1.47 -0.90±1.55 -1.64±2.91 -1.11±2.26 -1.31±2.70 -0.92±2.94 -0.51±2.40 -1.00±3.58 MaxProbdb5 -0.61±1.34 -0.66±1.49 -0.54±1.58 -2.27±2.92 -0.44±2.33 -0.91±2.71 -0.02±2.95 0.33±2.41 -0.35±3.62 Bostondb60 -1.03±1.50 -1.13±1.60 -0.86±1.65 -1.78±2.97 -1.50±2.38 -1.46±2.81 -0.83±2.99 -1.00±2.48 -1.24±3.71 Bostondb40 -1.10±1.44 -1.15±1.57 -0.88±1.59 -1.92±2.90 -1.52±2.36 -1.52±2.71 -0.86±2.96 -1.18±2.41 -1.30±3.74 Bostondb20 -0.20±1.42 -0.34±1.52 -0.02±1.60 -0.25±2.93 -0.54±2.31 -0.58±2.79 0.54±3.07 0.58±2.46 -0.11±3.78 Bostondb10 -0.66±1.46 -0.67±1.58 -0.47±1.62 -0.95±2.93 -0.71±2.36 -0.57±2.75 -0.34±2.99 -0.13±2.44 -0.51±3.74 Bostondb5 -0.71±1.40 -0.72±1.56 -0.58±1.60 -2.16±2.92 -1.00±2.38 -1.04±2.71 -0.38±3.03 0.02±2.44 -0.69±3.74 551 Table 3: Mean relative error ( 𝑉𝑇(𝐶𝑇)−𝑉𝑇(𝑝𝐶𝑇) 𝑉𝑇(𝐶𝑇) ) in VT between PET reconstructed with CT or 552 pseudo-CT for attenuation correction (mean ± SD in %). Bold font indicates the best-performing 553 variants used in subsequent analyses. 554 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 27 FIGURES 555 556 Figure 1: Schematic workflow for pseudo -CT (pCT) creation. The T1-weighted image of the 557 subject for which a pCT should be created is provided in its native space (individual space). 558 The database of CT and T1 -weighted MR image pairs is provided in MNI spac e. The 559 transformation of the T1-weighted image to MNI space is estimated first (A). UCL: a weighted 560 mean of the database CTs based on local similarity between database and subject T1-561 weighted images is computed and transformed back to individual space (B). Boston: A n 562 average of database CT images yields a template (C). mMaxProb: Following segmentation, a 563 tissue-selective average of database CT images yields a template (D). Templates are 564 transformed back to individual space (E). 565 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 28 566 Figure 2: CT and pseudo-CT for a representative subject. The complex air-bone interface in 567 the paranasal sinuses poses an issue in all pCT approaches. mMaxProb and UCL preserve 568 sharp tissue borders whereas Boston exhibits obvious smoothing across tissues. Clear 569 differences in size and tissue-bone composition are visible in the paranasal sinuses (1). 570 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint 29 571 Figure 3: Boxplot showing mean relative quantification error ( 𝑉𝑇(𝐶𝑇)−𝑉𝑇(𝑝𝐶𝑇) 𝑉𝑇(𝐶𝑇) ) between PET 572 reconstruction with CT and pCT. The three pCT methods UCL, Bostondb20 and MaxProb db5 573 represent the best-performing pCT approaches. Cerebellar grey matter (CGM). 574 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted December 24, 2025. ; https://doi.org/10.64898/2025.12.22.695942doi: bioRxiv preprint