Large scale comparison of maximum likelihood scatter scaling and tail-fitted scatter scaling in LAFOV PET/CT

Abstract

Background

Scatter scaling during the reconstruction of Positron Emission Tomography (PET) data is a crucial element for obtaining clinically applicable images with accurate quantification and high image quality. The current clinical standard for scatter scaling is fitting the tail regions of the single scatter simulation (SSS) estimate, which is termed Tail-Fitted Scatter Scaling (TFSS). This study aims to compare a Maximum Likelihood Scatter Scaling (MLSS) algorithm relative to TFSS using a NEMA IQ phantom investigation and a patient cohort including 500 patients using long axial Field-of-View (LAFOV) PET. The relative difference between the two scatter scaling algorithms was investigated using uptake values of 12 organs. Furthermore, the proximity of known regions showing high activity relative to the surrounding tissue was analysed.

Results

The NEMA image quality phantom study showed agreement between the expected activity concentration and the MLSS reconstructions. MLSS showed uptake values of 137.3 ± 3.4 kBq/mL in the largest sphere and 34.6 ± 0.5 kBq/mL in the background, closely matching the true concentrations of 136.6 kBq/mL and 35.0 kBq/mL, respectively. TFSS provided uptake values of 133.7 ± 3.5 kBq/mL in the largest sphere and 33.0 ± 0.9 kBq/mL in the background. MLSS showed higher uptake in the cold areas relative to TFSS. Mean recovery coefficients (RC_mean) showed that MLSS generally had coefficients closer to 1 relative to TFSS across the spheres of the phantom. The findings of the patient study showed a numeric relative difference below 2% when investigating organ uptake through the 12 organs.

Conclusion

MLSS provided results of high image quality comparable to the standard method of choice, TFSS, in the clinical routine. The phantom study showed that MLSS provided uptake values accurately relative to the known activity concentration, however less accurate within the cold sphere and insert. MLSS was found to provide robust results across a large patient cohort and is suggested as a suitable substitution for TFSS in the PET image reconstruction process.

Introduction

Positron Emission Tomography (PET) in combination with Computed Tomography (CT) provides physiological and anatomical information applicable for diagnosis, staging, and tumor quantification in oncology [1, 2]. The quality of the PET images is highly dependent on attenuation and scatter correction. Compton scattering is the dominant interaction for 511 keV photons in tissue [3]. The scattered events can, in principle, be rejected by energy window discrimination, but because energy losses are small and energy resolution of detectors limited the accepted remaining scatter fraction in a whole-body scan can be estimated to be up to 50% increasing with object (patient) size [4]. Scatter contributes to decreased image quality (reduced contrast) by misplacing the coincidence event location and causes quantitative errors within uptake estimation [5]. Correction factors for both attenuation and scatter correction are typically derived from CT images, which are acquired solely for these purposes, if a diagnostic CT is not necessary [6, 7]. The development of Long Axial Field-of-View (LAFOV) technology has increased the likelihood of detecting multi-scattered events, thereby amplifying scatter during PET acquisition [8, 9]. The impact of detected scatter is dependent on the tomograph energy resolution, acceptance energy window settings, and the tracer distribution [5, 10, 11]. The omission of scatter correction in the reconstruction causes a loss of resolution, contrast, and inaccuracies in quantification in the final image [12, 13].

Modeling of scatter scaling factors might become computationally demanding [14]. Many scatter correction algorithms use Single Scatter Simulation (SSS) which models scatter distribution [15]. A common approach to scatter scaling involves adjusting the simulated initial scatter estimate in the tail regions, under the assumption that only scatter is present in these areas. This method is known as Tail-Fitted Scatter Scaling (TFSS) [16]. TFSS is an iterative process that applies the scaling to regions outside the subject, as defined by the CT-based attenuation map mask. The final converged scatter scaling estimate from the iterative process is then used in the reconstruction [17]. It is known that TFSS possesses some performance limitations in cases where the tail regions have poor statistics, or if the tail regions are small, e.g. the subject occupy a large fraction of the FOV, or if the tails are incorrectly defined, e.g., patient movement during scan acquisition. Furthermore, artifacts such as photopenic areas, also known as halo artifacts, might be induced in areas with high organ-to-background uptake contrast due to even a small overestimation of the fraction being subtracted [18]. The halo artifact is prominently found near the excretion system, including the urinary bladder and the kidneys [19]. Other groups have suggested using Monte Carlo simulation or Deep Learning-based approaches to derive scatter scaling factors, aiming to overcome the performance limitations of TFSS [17, 20–22]. Deep Learning methods have been found to significantly reduce the computation time compared to traditional approaches [23] and are able to provide accurate results using LAFOV PET/CT [24]. As the energy resolution of PET scanners is improved, it is also suggested to return to energy-based trues estimation [25], which previously has been rejected as a method for scatter correction [5]. Some studies have showed positive results using energy-based scatter estimation as a substitute for TFSS [26, 27]. Another method for scatter scaling is Maximum Likelihood Scatter Scaling (MLSS), which was developed by Panin [14] and further improved by Rezaei et al. [28]. After each iteration of SSS the MLSS is utilized for obtaining the scatter scaling estimates, based on the entire emission image. Previous studies by Bal et al. (2021) have compared the MLSS algorithm to the TFSS algorithm using 71 2-[¹⁸F]fluoro-2-deoxy-D-glucose ([¹⁸F]FDG) and 11 68-Galium-DOTATE ([68Ga]-DOTATATE) patient studies using the Siemens Biograph mCT PET/CT system (Siemens Healthineers, Erlangen, Germany). The group found that the MLSS method showed no significant difference in diagnostic image quality, however, an additional phantom study showed that MLSS possessed the lowest bias, and thus may be preferred as standard method in the clinical routine due to robust performance [29]. Our group has previously demonstrated that MLSS reduces the appearance of halo artifacts significantly when it is used during the reconstruction of PET/MRI scans [30].

The scope of this study is to extend the investigations of Bal et al. [18] and Overbeck et al. [30], aiming to compare the MLSS algorithm relative to TFSS, in a large-scale study using high-sensitivity LAFOV [¹⁸F]FDG PET/CT scans.

Materials and methods

Phantom scan

A National Electrical Manufacturers Association (NEMA) IQ phantom was scanned using a LAFOV PET/CT system, Siemens Biograph Vision Quadra (Siemens Healthineers, Erlangen, Germany). The system was cross calibrated using ¹⁸F to a reference activity meter and a gamma counter. A total activity of 335.6 MBq was present in the phantom at scan start to study the differences of the method in a scenario with low noise. The spheres have internal diameters of 10, 13, 17, 22, 28 and 37 mm. 4.9 MBq was present in the spheres corresponding to the activity concentration of 136.6 kBq/mL [¹⁸F]FDG. The second largest sphere (28 mm) was filled with distilled water in the scope of visualizing a cold region. An activity of 330.7 MBq was added to the background of the phantom corresponding to 35.0 kBq/mL at scan start, aiming at a ratio between spheres and the background of 4:1. Samples (0.5 mL) were taken in triplicate from the phantom background and the beaker used for the filling of the spheres, and counted in the gamma counter.

A CT image was acquired using 200 mAs and 100 kVp. The PET scan was acquired in the duration of 10 min.

Patient cohort

500 patients (300 females, 200 males) from the clinical routine were retrospectively included in the large-scale study. The patients’ ages were in the range from 14 to 93 years (61.9 ± 14.9 years). The body mass index (BMI) of the patients was on average 25.5 ± 5.1 (range: 15.2–52.9). An activity of 3 MBq per kilo body weight (223.4 MBq ± 50.8 MBq) [¹⁸F]FDG was administered to the patients approximately one hour before scan acquisition. The scans were performed with the arms above their heads, if possible, using a LAFOV PET/CT system for a duration of five minutes. The retrospective study was approved by the institutional legal committee (no. p-2024-16833, approval date: 5/6-2024), thus patient consent was not needed. All patient data were treated fully anonymously by the European General Data Protection Regulation (GDPR).

Image reconstruction

The reconstruction utilized 3D Ordinary Poisson Ordered Subset Expectation Maximization (OP-OSEM) with 4 iterations, 5 subsets, point spread function (PSF) modelling, and a maximum ring difference of 322 [31]. A 2 mm Gaussian filtering were applied the reconstructed images. The final reconstructed images have a voxel size of 1.65 × 1.65 × 1.65 mm³. The scatter corrections methods were implemented as a MATLAB prototype (Siemens Healthineers, Erlangen, Germany) integrated with executables using an investigational software prototype (e7tools, Siemens Healthineers) providing the reconstructed images of the PET scan using either TFSS or MLSS as the method for scatter scaling.

As previously described, TFSS adjusts the SSS estimate by scaling the tail region outside the object. These regions are determined by a forward projection of the 3D image using linear attenuation coefficients derived from the CT. The algorithm iteratively estimates scaling factors used to correct the final reconstructed image [18]. The MLSS method begins with an uncorrected image for the first SSS iteration. Unlike TFSS, MLSS uses the entire emission sinogram for estimating the scatter scaling estimates. After each SSS iteration the scatter scaling factors are updated using a maximum likelihood approach, which improves the alignment between the measured and estimated data. A boundary mask defines the subject: voxels outside this region are set to zero. The mask is initially created using a CT-based bed-removal attenuation correction map. If voxel values outside the mask exceed the mean intensity value of the boundary voxels, the mask is extended to include the respective voxels [32].

Overall, MLSS is a nested, iterative method that estimates scatter scaling factors directly from the sinogram. It uses a TOF-based maximum likelihood activity update, constrained to prevent negative values [18]. The method is explained in detail by Bal et al. [18, 32].

During reconstruction, the following corrections were applied: bed removal, normalization, deadtime, randoms smoothing and subtraction, attenuation correction, decay correction, frame-length corrections, and whole-body scatter correction.

For simplicity, reconstructions using either MLSS or TFSS as scatter scaling methods are termed MLSS or TFSS reconstructions throughout this paper.

Quantitative image evaluation

The phantom scan was analysed by comparing the administrated and externally validated activity concentrations relative to the image-derived activity from the MLSS or TFSS reconstructed images. The image-derived activity was obtained based on a volume of interest (VOI) in the center of the spheres, the cold insert, and a VOI selected in the background of the phantom. Comparison of image-derived activities relative to the reference activity concentration was performed based on a mean recovery coefficient (RC_mean). RC_mean was the ratio of the mean activity concentration of the VOI in the spheres and the reference activity concentration. The spheres were termed S1-S6, of which S1 was the sphere with the largest diameter and S6 was the sphere with the smallest diameter.

Regional differences between the scatter scaling algorithms were investigated using the patient PET images by calculating the relative difference images using the TFSS reconstructed image values as reference: $\frac{(MLSS-TFSS)}{\mathit{TFSS}}\times 100\%$. The quantitative image evaluation of the patient cohort was performed using CT-based organ masks. The TotalSegmentator algorithm (version 1) was used to obtain the organ masks of the patient [33]. This included a mask of the following organs: spleen, kidneys, liver, aorta, lower lobe of the right lung, heart myocardium, left ventricle, brain, colon, and the urinary bladder. Furthermore, a mask of the right gluteus maximus was obtained in the scope of analysing a muscular structure in close proximity to the urinary bladder which presents with a high signal-to-background activity ratio.

A quantitative analysis of areas surrounding high-activity areas was performed. The urinary bladder and the kidneys were selected for this evaluation. Masks of the bladder and kidneys were dilated in the three dimensions leading to an increase in mask size corresponding to an enlargement of 16.5 mm in the axial planes, relative to the original size, and 20 mm in the transaxial plane. The original mask was then subtracted from the dilated mask resulting in an outline mask of the surroundings of the kidneys and the urinary bladder. The dilations of the selected masks and the extraction of uptake values within all obtained masks were performed using the McConnell Brain Image Centre software (MINC) followed by a comparative analysis performed using Python. All masks were used for extraction of the mean uptake concentration within the organs and the result was presented as the mean relative difference between the reconstructed image using either MLSS or TLSS. Patients showing an absolute relative difference larger than 10% in at least one organ were categorized as “outlier-patients”.

Results

Phantom study

The measured reference activity concentration of the spheres was 136.6 kBq/mL, and the background activity was 35.0 kBq/mL (ratio 3.9:1) based on extracted samples. Figure 1 visualises the reconstructed images of the phantom scan using both methods along with the VOIs of the spheres, the cold insert and the background. The image-derived mean activity concentration of S1 was 137.3 ± 3.4 kBq/mL and 133.7 ± 3.5 kBq/mL, for MLSS and TFSS, respectively, and the background VOI provided the values of 34.6 ± 0.5 kBq/mL and 33.0 ± 0.9 kBq/mL, respectively. Figure 2 visualizes the uptake value through a line passing the background, the cold insert, and S1. The figure shows that MLSS provided accurate uptake values in the hot areas relative to the dashed line representing the reference activity concentration. The TFSS reconstructed image shows an overcorrection of scatter when approaching the center of the phantom, whereas MLSS shows higher activity in the cold lung implant in the center of the NEMA IQ phantom. The activity concentration inside the insert using MLSS was 1.9 ± 0.3 kBq/mL, whereas TFSS showed uptake values of 0.7 ± 0.1 kBq/mL. The activity of the cold sphere was 5.6 ± 1.3 kBq/mL and 4.0 ± 1.1 kBq/mL for the MLSS and TFSS reconstruction, respectively. Thus, both are observed to provide too high uptake value, with MLSS having the largest difference to the expectation. Table 1 summarizes the extracted activity concentration relative to the extracted VOIs of the reconstructed images along with the RC_mean. RC_mean of the methods showed an approximate constant difference across the different spheres with MLSS having approximately 0.02 higher RC_mean values. The RC_mean of S1 showed a value above 1 for MLSS, whereas all other RC_mean values are below 1.

Fig. 1.

Visualization of the reconstructed image of the phantom study using either MLSS or TFSS. The volumes used for comparison of the reconstruction methods are illustrated along with the mean activity concentration inside the VOI

Fig. 2.

Visualization of the activity concentration through the indicated line of the phantom. The blue line represents the activity concentration of the MLSS reconstructed image, whereas the red line indicates the TFSS reconstructed image. The green dashed line represents the activity concentrations provided by the gamma counter

Table 1.

Overview of the PET-measured mean activity concentration in the spheres and the respective mean recovery coefficient (RC_mean) relative to the expected concentration of 136.6 kBq/mL

Sphere number	Sphere diameter	MLSS		TFSS
		Activity conc. [kBq/mL]	RCmean	Activity conc. [kBq/mL]	RCmean
1	37 mm	137.3 ± 3.4	1.01	133.7 ± 3.5	0.98
2	28 mm (cold)	5.6 ± 1.3	–	4.0 ± 1.1	–
3	22 mm	135.1 ± 10.7	0.99	132.0 ± 11.0	0.97
4	17 mm	134.3 ± 14.3	0.98	130.0 ± 14.9	0.95
5	13 mm	130.5 ± 14.9	0.96	126.9 ± 15.0	0.93
6	10 mm	103.2 ± 11.4	0.76	99.3 ± 11.3	0.73

Patient study

The investigation of the mean organ uptakes is shown in Fig. 3. The violin plots illustrate that most of the organs showed a relative difference between scatter correction methods in tracer uptake close to 0%. This was shown numerically in the listed findings of the mean values of the relative difference stated in Table 2. Based on the average across the patient cohort of the mean relative differences, it was found that MLSS showed slightly higher uptake values in eight out of twelve investigated organ masks. The remaining four; the brain, colon, right gluteus maximus, and the urinary bladder, showed slightly lower uptake values when MLSS was used. The largest numeric relative difference of -2% was found in the brain uptake, whereas all other organs showed an average deviation below 1%. The analysis of the organ uptake revealed a group of patients having a numeric relative difference above 10%. The organ showing the largest number of outlier patients was the lower right lung by the inclusion of 13 patients, which corresponded to 2.6% of the total cohort population. In total 26 patients were marked as outlier patients, due to a relative difference above 10% in at least one organ. Two patients showed an absolute relative difference above 10% in nine organ masks, seven patients showed a relative difference above 10% in 2–6 organs, and the remaining 13 patients showed outliers uptake values in a single organ. The BMI of the outlier patients was on average 28.9 ± 6.3, $\mu \pm \sigma$ (range: 20.7–49.0). The investigation of the uptake surrounding typical high uptake organs such as the urinary bladder and the kidneys showed that TFSS estimated higher uptake around the urinary bladder whereas MLSS estimated lower values around both kidneys (Fig. 4). The mean relative difference across all patients was found to be -3.6% ± 3.6% for the urinary bladder, and 0.5% ± 3.7% and 0.5% ± 3.6% for the left and right kidney, respectively.

Fig. 3.

Visualization of the distribution of the relative difference $(MLSS-TFSS)/TFSS\times 100\%)$ mean uptake values of the investigated organs in all patients. The figure illustrates that all examined organs have outliers showing a relative difference above 10% absolute relative difference

Table 2.

Summary showing the investigated organs and the mean relative difference in percent between the two scatter scaling methods of the respective organ

Organ	Average relative difference [%] ($\mu \pm \sigma$)	No. of patients with > 10% relative difference
Spleen	0.1 ± 2.7	10 (2.0%)
Right Kidney	0.2 ± 2.3	4 (0.8%)
Left Kidney	0.1 ± 2.1	2 (0.4%)
Liver	0.2 ± 2.7	9 (1.8%)
Aorta	0.4 ± $2.8$	6 (1.2%)
Lower Right Lung	0.6 ± 3.7	13 (2.6%)
Heart Myocardium	0.2 ± 2.7	9 (1.8%)
Left Ventricle	0.2 ± 2.5	7 (1.4%)
Brain	− 2.0 ± 1.7	1 (0.2%)
Colon	− 0.3 ± 2.9	9 (1.8%)
Right Gluteus Maximus	− 1.0 ± 2.2	3 (0.6%)
Urinary Bladder	− 0.9 ± 1.3	1 (0.2%)

Furthermore, overview of the number of outliers found with a numeric relative difference larger than 10% between the two methods of use. The occurrence of outliers is reported both as the actual number and in percent of the cohort

Fig. 4.

Illustration of the findings of the surrounding uptake around the urinary bladder, the left and the right kidney. The violin plots illustrate the distribution of the relative differences between the two scatter scaling methods

The violin plots of Figs. 3 and 4 showed that MLSS generally had slightly larger uptake values in the torso region relative to TFSS, whereas MLSS showed uptake values of lower magnitude in the head and extremities relative to TFSS. This was also observed when plotting the relative difference of the patient reconstructions. A visualization of a representative patient showing a minor relative difference in uptake values is presented in Fig. 5. The figure includes two transaxial slices illustrating that the relative difference in a reference organ such as the liver is close to zero and that the area just outside the bladder did not possess larger differences in uptake. Figure 6 illustrates a patient of which the reconstructed images indicate a relative difference between the methods of above 10% in nine organs. The transaxial views indicated that the large differences were mostly observed in low-activity areas that are very sensitive to relative measurements.

Fig. 5.

Illustration of the reconstructed images using either MLSS or TFSS and the relative difference given in percent of a patient showing minor differences when the two methods for scatter scaling is used. The patient has the BMI of 25.1 and the provided activity is 259 MBq

Fig. 6.

Illustration of the reconstructed images using either MLSS or TFSS and the relative difference given in percent of a patient with a relative difference above 10% in nine organs. The BMI of the patient is 31.4 and the provided activity is 305 MBq

Discussion

This study compared two scatter scaling methods: MLSS versus the current clinical standard, TFSS. The phantom study reflected image-derived activity concentrations that correspond to the expected externally measured values. MLSS showed the most accurate uptake values based on the calculated RC_mean values, which were approximately constant across the methods. Recovery coefficients calculated based on the maximum or peak activity concentrations would be more sensitive to noise, highlighting the advantage of using mean uptake values, that however is influence by the choice of VOI. Overall, the RC_mean of both methods showed agreement relative to the reference measurement, but MLSS yielded uptake values closer to the reference measurement. Investigation of cold regions showed that both methods provided higher uptake values for the cold sphere relative to the insert. The difference of the size and the material of the insert and the presence of water in the sphere led to higher uptake values in the sphere relative to the cold insert. TFSS showed the lowest uptake values in both regions, thus providing the most accurate image-derived uptake. The similarity of the phantom image sets contributed to the understanding that MLSS provides reconstructed images of clinical standards and might be more accurate than the current standard due to the RC_mean values.

The presented results of the reconstructed images from the patient cohort showed that the uptake values in the investigated organs had minor relative differences (Table 2). The greatest limitation was that the “true” uptake values of a patient are unknown. The magnitude of the relative differences was of minor size below the natural bias of the reconstruction process [18]. The comparison showed an overall agreement between the two scatter scaling methods. As the liver is used as a reference organ during daily clinical routine, the agreement in uptake within this organ was considered a strong indicator of the methods performing in accordance with each other. The inspection of the relative difference maps comparing the scatter scaling methods visually agreed with this interpretation. The examples of the difference maps shown in Figs. 5 and 6 showed that MLSS generally had marginally higher values in the torso region, whereas TFSS showed larger uptake values in the head and extremities. This may be due to the application of TFSS based scatter scaling factors across the torso including areas of high uptake which may lead to inaccuracies in the fitting process. It was investigated whether this was caused by PET and CT mismatch due to patient movement between acquisitions. The inspection concluded that there was no clear visible movement between acquisitions, however, minor motion may have caused activity to be misclassified into the tail region. TFSS is more sensitive to motion than MLSS, as the latter incorporates the entire dataset. The observation of higher uptake values in the lungs within MLSS reconstructed images of the patients agreed to the findings of the phantom study, where MLSS showed higher uptake values in the cold regions. Thus, MLSS shows higher activity concentrations in low-count areas, which were less accurate than the image derived results from the TFSS reconstructions. It is commonly known that TFSS may show a decrease in performance for the reconstruction of patient examinations where the patient occupies a greater part of the axial FOV [32]. The BMI of the full patient cohort is described as 25.5 ± 5.1 whereas the patient sub-group termed as outlier patients had an average BMI of 28.9 ± 6.3 indicating that the small number of patients where the comparison of the methods showed larger differences primarily included patients that were categorized as overweight.

Another performance limitation found when TFSS is used during reconstructions is the so-called halo artifact. This artifact could have a major impact on clinical decision-making. Due to the halo artifact being mostly observed in regions with high uptake contrast, the dilation of the urinary bladder, and kidney masks were investigated. The presented results contradicted the expectation by showing generally larger uptake values within the TFSS reconstructed images around the bladder. Halo artifacts were, however, not observed in the reconstructed images using either method within this patient cohort. This may be due to the general improvement of the LAFOV PET scanner software.

Alternative methods for scatter scaling of PET images are suggested to be performed using deep-learning or Monte Carlo-based approaches [34, 35]. The study by Li et al. [23] utilized so-called histo images generated based on TOF information and a mapping of the maximum likelihood annihilation position. The group concluded that the method was quantitatively accurate and approximately three times faster than conventional scatter scaling algorithms. The study by Yang et al. [6] demonstrated that deep learning could potentially benefit both scatter and attenuation correction, enabling CT-less procedures of the patients. However, the group also highlighted that the deep learning-based approach may miss lesions, followed by a conclusion stating that more work was needed to use deep learning-based scatter scaling algorithms. Even though other methods such as deep learning-based method are under development, the reconstruction tool using MLSS is found to be stable and robust across a large-scale cohort. The findings of this study presented MLSS as a proper substitute for scatter scaling during the reconstruction of PET images based on the resulting images comparable to clinical standards, and thus, it is suggested to obtain even more accurate uptake values based on the findings of the phantom study showing the accurate activity concentrations when MLSS was used. The agreement between MLSS and TFSS in this study was in line with the results of the study by Bal et al. [18].

Future investigations include inspecting the performance of MLSS compared to the clinical standard TFSS using other tracers, such as [⁶⁸Ga]Ga-PSMA, due to the known higher frequency of halo artifacts and concerning the investigation of the difference of the methods using a tracer with high specificity and the added complication of prompt gamma photons. Other fields of interest regarding quantification of PET reconstructed images using MLSS could be by comparison of simulated data and using studies of low-count examinations.

Conclusion

The study presented that MLSS is suitable for scatter scaling in PET images and the use of MLSS is comparable to the current standard method of choice in LAFOV [¹⁸F]FDG PET/CT clinical applications. Furthermore, the study presented that MLSS provided the most accurate uptake values based on a phantom study, however showed less accurate uptake values in the cold regions. Based on this work, it is concluded that MLSS is a robust method that serves as a proper substitute for TFSS.

Author contributions

Conceptualization, T.L.A., and F.L.A.; methodology, N.O., T.L.A., and F.L.A.; software, N.O., M.T., and M.C.; validation, N.O.; formal analysis, N.O.; investigation, N.O.; resources, F.L.A.; data curation, N.O, and S.H..; writing—original draft preparation, N.O.; writing—review and editing, N.O., S.H., M.T., M.C., T.L.A., and F.L.A.; visualization, N.O.; supervision, T.L.A., and F.L.A.; project administration, T.L.A., and F.L.A.; All authors have read and agreed to the published version of the manuscript. M.C. and M.T. are full time employees of Siemens Medical Solutions Inc.

Funding

Open access funding provided by Copenhagen University. This project has received no funding.

Availability of data and materials

Data supporting reported results can be obtained via contact with the corresponding author upon reasonable request and legal approval. The data are not publicly available due to no public data sharing agreement.

Declarations

Ethics approval

The procedures performed in the study were in according with the ethical standards of the institutional legal committee. This retrospective study was approved by the institutional legal committee (no. p-2024-16833, approval date: 5/6-2024).

Consent to participate

Human ethics and consent to participate declarations is not applicable.

Competing interests

Mohammadreza Teimoorisichani and Maurizio Conti are employed by Siemens Medical Solutions Inc. Otherwise, the authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.