Spatial non-uniformity of liver virtual non-contrast attenuation on dual-layer spectral CT: Impact of body habitus and peripheral right-lobe error

Abstract

Purpose

To evaluate the global and spatial accuracy of virtual non-contrast (VNC) liver attenuation derived from dual-layer spectral CT and its association with patient body habitus.

Methods

This retrospective single-center study included 99 patients examined with true non-contrast (TNC) and three post-contrast phases on dual-layer CT. Global TNC–VNC agreement was assessed in liver segments S2, S7, and S8. Detailed center-to-periphery analysis in segments S7 and S8 was performed in 87 patients. Spatial non-uniformity was quantified using the spatial error gradient (SEG), defined as the difference between peripheral and central absolute VNC error.

Results

Median absolute whole-liver TNC–VNC differences were 3.0 HU in the arterial phase, 4.4 HU in the portal venous phase, and 3.1 HU in the delayed phase. Peripheral VNC error exceeded central error in all phases; the largest difference was observed in the portal venous phase (4.4 vs 2.3 HU), with a corresponding SEG of 2.4 HU (p < 0.001). Abdominal effective diameter showed a weak but statistically significant correlation with SEG in the portal venous and delayed phases (both p = 0.004). In an exploratory ROC analysis, effective diameter predicted elevated peripheral error with an AUC of 0.746.

Conclusion

Dual-layer CT-derived VNC images showed good overall agreement with TNC, but exhibited a population-level tendency toward spatial non-uniformity, with peripheral error exceeding central error. Based on these exploratory data, in evaluation of diffuse liver attenuation, central liver measurements may be preferred over peripheral measurements in larger patients.

Highlights

• •Dual-layer VNC matched TNC well for global liver attenuation
• •VNC error was greater at the peripheral right liver lobe
• •Spatial error rose with increasing abdominal effective diameter
• •Central liver ROI were more reliable than subcapsular ROI
• •Repeat SEG measurements showed poor-to-moderate intra-reader agreement

.

1. Introduction

Spectral computed tomography (CT), particularly dual-layer spectral CT (DLCT), has significantly expanded the capabilities of abdominal radiology through virtual non-contrast (VNC) reconstructions [1], [2]. Generating VNC images from post-contrast acquisitions reduces cumulative radiation dose and optimizes scanner workflow by eliminating true non-contrast (TNC) phases [1], [2], [3]. An intrinsic advantage of detector-based spectral separation (DLCT) is that VNC data are generated routinely without requiring pre-scan protocol modifications [1], [2]. The accuracy of these reconstructions relies heavily on the quality of the material decomposition algorithm used to separate iodine from soft tissue [1], [4].

While VNC images demonstrate excellent global agreement with TNC acquisitions in central abdominal regions, regional deviations and edge-of-field artifacts have been reported [4], [5], [6], [7], [8], [9], [10]. These discrepancies may carry practical significance when evaluating peripheral parenchyma in larger solid organs, such as the liver. Inaccuracies in measured attenuation can stem from geometric artifacts, beam hardening, and photon starvation caused by adjacent osseous structures like the rib cage [4], [9], [10]. From a clinical standpoint, these phenomena are particularly relevant for opportunistic quantitative liver attenuation assessment in diffuse hepatopathies [11], [12], [13], [14], [15].

However, existing literature has predominantly evaluated the liver as a single, homogeneous volume, relying on averaged global regions of interest (ROIs) [5], [6], [7]. Whether peripheral deviations represent localized geometric artifacts or signify a systemic failure of spectral separation in larger patients remains unquantified. Therefore, this study aimed to evaluate global and regional TNC–VNC agreement in DLCT and to quantify spatial non-uniformity using a novel parameter: the Spatial Error Gradient (SEG).

2. Materials and methods

2.1. Patient cohort

This retrospective, single-center study was approved by the Ethics Committee of the University Hospital Brno. Given the retrospective nature of the study, informed consent was waived. In our study only patients examined on a DLCT (using a dedicated liver protocol (which included 3 standard postcontrast phases)) between January 2020 and October 2024 were included. Exclusion criteria were: 1. severe cirrhosis with substantial liver lobe atrophy 2. liver lesions bigger that 2.0 cm and any postoperative changes preventing correct ROI placement 3. portal vein thrombosis and cavernous transformation 4. known primary or secondary malignancy.

2.2. Image acquisition and reconstruction

All examinations were performed on a DLCT scanner (IQon, Philips Healthcare). The standard acquisition parameters were as follows: tube voltage 120 kVp, pitch factor 0.89, and collimation 64 × 0.625 mm. Standard TNC and conventional post-contrast images were reconstructed using iterative reconstruction (iDose, level 5). Images were reconstructed with a slice thickness of 2.0 mm and an increment of 1.0 mm. A non-ionic iodinated contrast agent (Iomeron 350 mg I/mL, Bracco, Italy; or Optirey 350, Guerbet, France) was administered intravenously via a power injector at a flow rate of 3.5–4.0 mL/s. The contrast volume was weight-adapted (1.2 mL/kg body weight), followed by a 40 mL saline chaser. Image acquisition for the post-contrast phases was timed using a bolus-tracking technique with a threshold of 150 HU in the descending aorta. The arterial, portal venous, and delayed phases were acquired at 5 s (or 12 s for pancreatic phase), 45 s, and 300 s after reaching the trigger threshold, respectively. Spectral VNC images were generated (using spectral reconstruction level 5) from the post-contrast spectral data using the manufacturer's proprietary software (IntelliSpace Portal, Philips, version 15.0). For the evaluation of the images, a standard abdominal window setting (window width 360 HU, window center 60 HU) was used.

2.3. Quantitative image analysis

To assess global VNC accuracy, parenchymal ROIs of the size 5–6 cm2 were placed in liver segments S2, S7, and S8, avoiding focal lesions and large vessels based on the portal venous phase. These ROIs were copied to TNC and all VNC phases, with manual adjustments made when necessary to correct for minor inter-phase respiratory motion (Fig. 1). The corrections were necessary in approximately 20% of patients.

Fig. 1.

Global liver parenchymal attenuation measurements. Illustrative example of ROI placement for global liver parenchymal attenuation analysis. ROIs were initially placed in the portal venous image in segments S2, S7, and S8, avoiding focal lesions and large vessels. The ROIs were then transferred to the true non-contrast (TNC) image and to virtual non-contrast (VNC) reconstructions from all post-contrast phases, with manual adjustment when necessary to ensure anatomical correspondence.

For quality control of spectral separation, attenuation was also measured in non-enhancing reference tissues: the right psoas major muscle, subcutaneous abdominal fat, and extracorporeal background air (size of each ROI was 2.5–3.5 cm2).

(Fig. 2, Fig. 3).

Fig. 2.

Quality control and verification of stability of spectral separation – muscle. Illustrative example of ROI placement in the right psoas major muscle for quality control of spectral separation. The ROI was initially placed in the portal venous image and then transferred to the TNC image and to VNC reconstructions from all post-contrast phases, with manual adjustment when necessary.

Fig. 3.

Quality control and verification of stability of spectral separation – fat, air. Illustrative example of ROI placement in the subcutaneous fat of the anterior abdominal wall and in background air outside the patient’s body for quality control of spectral separation. The ROIs were initially placed on the portal venous image and then transferred to the TNC image and to VNC reconstructions from all post-contrast phases, with manual adjustment when necessary.

To evaluate segmental accuracy and spatial non-uniformity, a dedicated center-to-periphery analysis was conducted in the right lobe (segments S7 and S8). Central ROIs were placed in the inner third of the parenchyma near the inferior vena cava. Peripheral ROIs were placed in the outer third of parenchymal depth, maintaining a minimum 5 mm margin from the capsule and extrahepatic structures whenever anatomically feasible (Fig. 4). All measurements were performed by a board-certified radiologist. To evaluate methodology reproducibility, inter- and intra-reader reliability were tested in a randomly selected subset of 50 patients. All global and center-to-periphery measurements were independently repeated by a second blinded reader (a radiology resident with 2 years of CT experience). Intra-reader repeatability was evaluated by the primary reader four months later on the same subset (n = 50).

Fig. 4.

The segmental accuracy of VNC in the liver parenchyma. Illustrative example of dedicated ROI placement for detailed segmental analysis of VNC accuracy. Central and peripheral ROIs were placed in segments S7 and S8 on the portal venous image, avoiding focal lesions and large vessels, and then transferred to the TNC image and to VNC reconstructions from all post-contrast phases, with manual adjustment when necessary. The difference between peripheral and central absolute VNC error was defined as the spatial error gradient (SEG).

Abdominal effective diameter (ED) was calculated at the Th12/L1 axial level from maximum anteroposterior (AP) and laterolateral (LL) dimensions:

$$\mathit{ED}=\sqrt{\mathit{AP}*\mathit{LL}}$$

To objectively evaluate the spatial degradation of the image, the following absolute errors relative to the reference image (TNC) were defined:

Central VNC error (ΔVNC_center): |HUTNC_center - HUVNC_ center|.

Peripheral VNC error (ΔVNC_periphery): |HUTNC_ periphery - HUVNC_ periphery|.

Spatial error gradient (SEG): calculated as ΔVNC_periphery - ΔVNC_center.

2.4. Outcome measures

The primary endpoint of the study was the evaluation of spatial error gradient (SEG) in the portal venous phase. Secondary endpoints included arterial and delayed phase SEG, correlation between SEG and ED, and global/segmental TNC–VNC agreement. All secondary threshold and correlation analyses were designated as exploratory.

2.5. Statistical analysis

Statistical analyses were performed using SPSS (v28.0, IBM Corp.,Armonk, NY, USA). Continuous variables are expressed as medians and interquartile ranges (IQR). Normality was evaluated using the Shapiro–Wilk test. Agreement between TNC and VNC attenuation was assessed using Bland–Altman analysis (mean bias, 95% limits of agreement [LoA], and proportional bias checks). An exploratory attenuation-based reclassification analysis evaluated whether VNC error altered patient assignment using a < 40 HU threshold (indicative of at least moderate steatosis). Discordance rates were evaluated across whole-liver, central, and peripheral regions of the right liver lobe using Cohen’s kappa and McNemar’s test. The direction of discordance was classified as VNC overclassification (VNC <40 HU with TNC ≥40 HU) or VNC underclassification (VNC ≥40 HU with TNC <40 HU).

Relationships between ED and SEG (and psoas muscle controls) were assessed using Spearman’s rank correlation coefficient. Receiver operating characteristic (ROC) analysis evaluated the discriminatory ability of ED to predict a high peripheral error, defined as absolute deviation at the periphery in the portal venous phase exceeding 10 HU (ΔVNCpv_periphery > 10 HU). Optimal cut-offs were determined via the Youden index. To evaluate the robustness of the ROC-derived threshold across different clinical error definitions, a sensitivity analysis was also performed for peripheral error thresholds of 5 HU, 8 HU and 15 HU.

Inter- and intra-reader reliability were evaluated using intraclass correlation coefficients (ICC). Two-way mixed-effects absolute-agreement models were used for intra-reader analyses, while two-way random-effects absolute-agreement models were used for inter-reader analyses. To control for multiple comparisons across secondary endpoints, p-values were adjusted using the Holm–Bonferroni method. Adjusted p < 0.05 was considered statistically significant.

3. Results

3.1. Patients’ demography

From 1004 screened archive records, 905 patients were excluded based on predefined criteria (Fig. 5). The primary global TNC–VNC analysis included 99 patients (70% male; median age 72 years, range 32–85). For the detailed center-to-periphery analysis, 12 additional patients were excluded due to small liver volume or anatomical constraints preventing dual ROI placement, leaving 87 patients (69% male; median age 73 years, range 32–85). The reproducibility subset with 50 randomly selected patients was used for all inter- and intra-reader evaluations (Fig. 6).

Fig. 5.

Study flowchart. Flowchart showing patient selection, reasons for exclusion, and the final cohorts included in the primary TNC/VNC analysis and in the dedicated center-to-periphery analysis.

Fig. 6.

Study flowchart. Flowchart summarizing the overlap between the primary agreement cohort, SEG cohort, inter-reader subset, and intra-reader subset.

3.2. Spectral separation stability in reference tissues

Analysis of reference non-enhancing tissues confirmed global algorithmic stability. Bland–Altman analysis showed minimal mean bias for the psoas major muscle (−1.04HU in the portal venous phase) and background air, though 95% LoA remained broad across phases. Subcutaneous fat exhibited a slightly higher systematic VNC overestimation, consistent with peripheral field-of-view attenuation behavior (Table 1).

Table 1.

Measurement agreement between true non-contrast and virtual non-contrast images in reference non-enhancing regions.

Evaluated Region	Enhancement Phase	Mean Bias (HU)	95% Limits of Agreement (HU)	Proportional Bias (FDR p-value)
Muscle (Psoas major)	Arterial	−2.33	−10.79–6.13	0.047
	Portal Venous	−1.20	−10.07–7.67	0.002
	Delayed	−1.58	−10.65–7.50	< 0.001
Subcutaneous Fat	Arterial	6.33	−0.26–12.93	1.000
	Portal Venous	7.78	0.88–14.67	1.000
	Delayed	9.45	1.20–17.71	1.000
Background (Air)	Arterial	−0.79	−9.66–8.09	0.001
	Portal Venous	−0.84	−9.52–7.84	< 0.001
	Delayed	−1.11	−11.13–8.91	< 0.001

3.3. Global agreement and measurement reproducibility

Global TNC–VNC comparison demonstrated small overall mean biases across the whole liver (e.g.,+3.43 HU in the portal venous phase), but 95% LoA frequently exceeded ±10 HU threshold (Table 2).

Table 2.

Global and segmental measurement agreement between TNC and VNC liver attenuation.

Liver segment / zone	Enhancement Phase	Mean Bias (HU)	95% Limits of Agreement (HU)	Proportional Bias (FDR p-value)
Global Liver (S2 +S7 +S8)	Arterial	2.16	−4.72–9.04	0.019
	Portal Venous	3.43	−4.73–11.58	0.002
	Delayed	2.56	−5.52–10.64	0.004
Segment S2 (ventral)	Arterial	0.43	−9.77–10.63	< 0.001
	Portal Venous	1.03	−10.44–12.49	0.003
	Delayed	−0.39	−14.03–13.25	0.110
Segment S7 (dorsal/subcapsular)	Arterial	1.02	−9.40–11.43	0.002
	Portal Venous	2.38	−10.43–15.18	0.002
	Delayed	1.98	−9.93–13.89	< 0.001
Segment S8 (dorsal/subcapsular)	Arterial	1.51	−7.73–10.75	0.035
	Portal Venous	2.83	−8.05–13.72	0.039
	Delayed	1.89	−8.20–11.97	0.002

Standard segmental liver parenchymal measurements showed substantial-to-excellent inter-reader agreement (ICC range: 0.844–0.931; p < 0.001) with a mean absolute difference < 1.4 HU. Intra-reader agreement for center-to-periphery measurements demonstrated near-perfect reliability (ICC range: 0.960–0.990; p < 0.001) with absolute differences < 0.5 HU (Supplementary Tables 1–2).

3.4. Spatial VNC degradation and SEG

Center-to-periphery analysis revealed that VNC accuracy was spatially non-uniform in segments S7 and S8. Central VNC attenuation closely mirrored TNC values, whereas peripheral absolute error (ΔVNC_periphery) consistently exceeded central error (ΔVNC_center) across all phases, peaking in the portal venous phase (4.4 [IQR: 1.9–9] HU vs 2.3 [IQR: 0.8–4.1] HU, p < 0.001).

Median SEG was positive across all phases, reflecting a population-level tendency toward higher peripheral error. In the portal venous phase, a positive SEG (>0 HU) occurred in 60% of patients, with 49% exceeding 3HU (Table 3). However, inter-reader reproducibility for the derived SEG metric was poor (ICC = 0.210 in the portal venous phase) due to error propagation from dual ROI positioning between readers (Supplementary Table 1).

Table 3.

Spatial degradation of virtual non-contrast accuracy in the right liver lobe.

Imaging phase	Absolute central error (HU)	Absolute peripheral error (HU)	SEG parameter (HU)	p-value
Arterial	1.9 (0.9 – 3.4)	3.3 (1.5 – 5.9)	1.4 (−0.4 – 4.0)	0,01
Portal venous	2.3 (0.8 – 4.1)	4.4 (1.9 – 9.0)	2.4 (−0.3 – 5.6)	< 0.001
Delayed	1.9 (0.6 – 4.1)	3.9 (1.7 – 8.6)	1.8 (−0.1 – 4.7)	< 0.001

3.5. Impact of body habitus on spatial error

In the center-to-periphery cohort (n = 87), median abdominal ED was 30.9 cm (range 24.6–39.3 cm). Spearman analysis revealed a weak positive correlation between ED and SEG in the portal venous (rs = 0.305; p = 0.004) and delayed phases (rs = 0.303; p = 0.004). In contrast, psoas muscle control attenuation showed no correlation with ED across any phase (p = 0.353–0.945) (Table 4).

Table 4.

Correlation between patient body habitus and spatial image degradation.

Evaluated parameter	Imaging phase	Correlation with patient thickness (rs)	p-value
Hepatic spatial error gradient (SEG) (subcapsular artifact)	Arterial	0.099	0.360
	Portal venous	0.305	0.004
	Delayed	0.303	0.004
Absolute deviation in muscle (internal central null standard)	Arterial	0.101	0.353
	Portal venous	−0.008	0.945
	Delayed	−0.022	0.841

3.6. High-error thresholds and reclassification

In the portal venous phase, 18% of patients (n = 16) exhibited a high peripheral error (ΔVNCpv_periphery > 10 HU). Peripherally, 70% of VNC values were higher than TNC, while 30% were lower. Exploratory ROC analysis indicated moderate discriminatory power for ED in predicting peripheral error > 10 HU (AUC = 0.746; 95% confidence interval: 0.625–0.866; p < 0.001). An optimal exploratory ED cut-off of 32.4 cm yielded 81.3% sensitivity (95% CI: 54.4–96.0%) and 76.1% specificity (95% CI: 64.5–85.4%). To evaluate the stability of this 32.4 cm cut-off, the sensitivity for detecting peripheral errors exceeding 5 HU and 8 HU was 54.5% and 80.0%, respectively. No patient in our cohort exhibited a peripheral VNC error exceeding 15 HU.

Exploratory attenuation-based reclassification (<40 HU threshold) showed global VNC discordance in 10.1% (10/99) of patients (κ = 0.59; McNemar p = 0.021), primarily driven by VNC overestimation masking low attenuation. Region-matched analysis confirmed a spatial effect: peripheral measurements demonstrated a significantly higher discordance rate (13%; κ = 0.58; McNemar p = 0.012) than central measurements (6%; κ = 0.77; McNemar p = 0.375).

4. Discussion

Spectral CT scanners capable of generating virtual non-contrast (VNC) images have become an integral component of routine abdominal imaging, offering substantial benefits in radiation dose reduction and workflow optimization by eliminating true non-contrast (TNC) acquisitions [1], [2], [3]. However, the clinical viability of VNC reconstructions hinges on their technical accuracy and fidelity compared to conventional TNC acquisitions [5], [6], [7], [16].

Our study demonstrates that dual-layer spectral CT (DLCT) achieves robust global accuracy for quantitative liver parenchyma attenuation, aligning with prior literature that reported minimal overall measurement biases on the order of single Hounsfield units [5], [6], [7], [14], [15], [16]. Evaluating internal non-enhancing reference tissues (psoas major muscle and background air) revealed minimal mean bias, confirming that detector-based material decomposition does not suffer from a systemic, global failure. However, while global agreement across averaged parenchymal ROIs was acceptable, dedicated topographic analysis of the right liver lobe uncovered significant individual-level variability characterized by spatial non-uniformity. Specifically, absolute attenuation errors increased progressively from the deep central parenchyma toward the subcapsular periphery of segments S7 and S8.

Methodologically, previous studies investigating VNC accuracy predominantly relied on averaged density values from centrally placed ROIs in homogeneous parenchyma, effectively masking regional variance [5], [6], [7], [14], [15], [16]. Our findings do not contradict these established global data, but they add a critical spatial context. Interestingly, when evaluating segment S2, no significant difference between TNC and VNC was observed (unlike in S7 and S8). However, S2 cannot serve as a direct central reference for right-lobe subcapsular regions because it inherently differs from S7 and S8 in anatomical depth, adjacent tissue surroundings, and X-ray beam path lengths.

The spatial degradation observed toward the periphery of segments S7 and S8 correlated weakly with abdominal effective diameter in the portal venous and delayed phases (rs approx 0.30), but was not statistically significant in the arterial phase (rs = 0.099, p = 0.360). This phase dependence can be explained by the underlying mechanism of spectral material decomposition. In the arterial phase, parenchymal iodine accumulation is usually minimal, so the post-contrast liver density closely resembles the true non-contrast state. Because the algorithm performs only minor iodine subtraction, any attenuation-dependent spectral error is probably scaled down and clinically negligible [4], [8]. Conversely, during the portal venous and delayed phases, parenchymal iodine concentration is substantially higher and the resulting aggressive spectral subtraction could amplify subtle habitus-related decomposition errors, making the spatial error gradient statistically detectable [4], [8], [9], [10].

Nevertheless, because intra-ROI noise (standard deviation) and rib distance were not prospectively modeled, beam hardening and photon starvation near the bony rib cage remain hypothesized mechanisms [4], [9], [10]. Elevated peripheral image noise, lobar path-length variations, and subtle subcapsular partial-volume effects likely represent compounding or alternative factors influencing the Spatial Error Gradient (SEG).

Regarding measurement reproducibility, standard segmental parenchymal ROIs showed substantial-to-excellent inter-reader agreement (ICC 0.844) between an experienced radiologist and a radiology resident, confirming that general segmental sampling is robust and independent of reader experience. Intra-reader repeatability for center-to-periphery measurements was likewise near-perfect (ICC 0.960). Conversely, the derived SEG metric exhibited poor inter-reader reproducibility (ICC = 0.210). Mathematically, calculating a difference between two independent attenuation values propagates and amplifies subtle positioning discrepancies that would be negligible in single-ROI analysis. Consequently, SEG should be interpreted cautiously as an exploratory population-level marker of spatial non-uniformity rather than a robust metric for individual patient profiling.

From a practical perspective, our exploratory reclassification analysis provides insight into the potential clinical impact of VNC errors during opportunistic liver attenuation grading. Using a < 40 HU threshold, sampled whole-liver VNC attenuation resulted in a 10.1% category discordance relative to TNC, predominantly driven by VNC overestimation masking low attenuation. Region-matched analysis confirmed the spatial effect: peripheral right-lobe measurements yielded a 13% discordance rate compared to 6% centrally. Because no independent gold standard (such as MRI-PDFF or histology) was available, these findings reflect attenuation-category discordance rather than proven diagnostic steatosis misclassification [12], [13].

Crucially, the practical implication—preferring central parenchymal ROIs in larger patients (ED > 32.4 cm)—must be restricted strictly to the quantitative evaluation of diffuse liver diseases. This recommendation cannot be generalized to focal or heterogeneous steatosis, where subcapsular fat deposition is common and central sampling alone would be misleading. In complex or non-uniform parenchymal patterns, whole-liver volumetric VNC segmentation [15] or multi-region sampling represents a significantly more robust approach than single subcapsular ROIs.

This study has several limitations. First, its retrospective single-center design utilizing a single DLCT platform limits generalizability across different spectral CT technologies (e.g., dual-source or rapid kVp-switching systems). Second, minor inter-phase respiratory motion between TNC and post-contrast acquisitions may have introduced subcapsular registration errors despite visual alignment. Third, intra-ROI noise levels were not prospectively archived across ED strata to definitively isolate photon starvation from partial-volume effects. Fourth, the primary outcome relied on a derived metric (SEG) susceptible to error propagation, underscoring the need for semi-automated or volumetric segmentation in future multi-reader designs. Finally, the 32.4 cm ED threshold was derived from a cohort with a limited number of high-error events (n = 16), rendering it strictly hypothesis-generating and requiring prospective external validation.

5. Conclusion

Dual-layer spectral CT-derived virtual non-contrast images demonstrate good systematic agreement with true non-contrast liver attenuation, but exhibit localized spatial non-uniformity at the individual patient level. VNC attenuation errors are greater at the subcapsular periphery of the right liver lobe and show a weak association with patient body habitus. For quantitative assessment of diffuse liver attenuation on VNC images, central parenchymal measurements should be preferred over peripheral subcapsular regions, particularly in larger patients.

CRediT authorship contribution statement

Valek Vlastimil Jr: Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Tomas Andrasina: Writing – review & editing, Supervision, Methodology, Conceptualization. Tomas Rohan: Writing – review & editing, Validation, Methodology, Data curation. Katerina Pecinkova: Investigation, Formal analysis.

Ethics declaration

The authors have NOT obtained informed consent from participants or their legal representatives. Given the retrospective nature of the study, the requirement for written informed consent was waived.

This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. This study was approved by the Institutional Review Board. (Approval No. 24-080323/EK).

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work, the authors used ChatGPT primarily for language revision and for refinement of the manuscript’s form and readability. After using this tool, the authors carefully reviewed and edited the content as needed and take full responsibility for the content of the published article.

Funding

This work was supported by the Ministry of Health, Czech Republic – conceptual development of research organizations (FNBr, 65269705) and Ministry of Education, Youth and Sports (MUNI/A/1823/2025) of the Czech Republic.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

Supplementary material

Supplementary material

Data availability

The data underlying this article are not publicly available because they contain potentially identifiable patient information. De-identified data may be made available from the corresponding author on reasonable request and subject to institutional and ethics approval.