Photon-counting CT for paraspinal muscle fat quantification compared with MRI proton density fat fraction

Abstract

Background

To evaluate photon-counting CT (PCCT) derived 70 keV attenuation values and virtual noncontrast fat fraction (VNC FF) in quantifying paraspinal muscle fat infiltration, using MRI proton density fat fraction (PDFF) as the reference standard.

Methods

In this prospective study, 76 adults with low back pain underwent same-day lumbar PCCT and 6-echo q-Dixon MRI within a 2-hour interval. The cohort consisted of 76 participants (38 men, 38 women), with a mean age of 47.7 ± 14.0 years and a mean body mass index (BMI) of 24.6 ± 3.3 kg/m². VNC FF represents a material decomposition–based fat fraction obtained from PCCT. Regions of interest (ROI) were bilaterally drawn in the multifidus, erector spinae, and psoas major at four intervertebral disc levels (L2/3–L5/S1). Correlation analysis, linear mixed-effects regression, Bland–Altman analysis, and receiver operating characteristic analysis were performed.

Results

The 70 keV CT values showed a strong correlation with MRI PDFF at the ROI level (r = − 0.931), outperforming VNC FF (r = 0.876). At the subject level, correlations were consistently strong across intervertebral disc levels (r range, − 0.964 to − 0.975 for CT values; 0.766 to 0.896 for VNC FF) and muscle groups (r range, − 0.881 to − 0.984 for CT values; 0.827 to 0.966 for VNC FF). Regression modeling enabled derivation of an internally calibrated CT fat fraction (CTFF), which closely approximated MRI PDFF within the study cohort. Agreement in categorical fat infiltration grading (< 10%, 10–30%, 30–50%, > 50%) was moderate (κ = 0.623). For binary classification at the 30% threshold, CTFF demonstrated excellent diagnostic performance (AUC = 0.993, 95% CI: 0.990–0.997).

Conclusion

PCCT-derived 70 keV CT values showed strong agreement with MRI PDFF and enabled internal regression–based estimation of paraspinal muscle fat, with excellent diagnostic performance for fat infiltration classification.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12880-026-02497-8.

Introduction

Quantitative assessment of skeletal muscle composition has attracted increasing attention because of its association with functional decline and its relevance across a range of clinical conditions [1, 2]. Loss of muscle mass and deterioration in muscle quality often coexist; the latter is primarily characterized by fatty infiltration and structural alterations [3–5]. These changes are associated with impaired physical function and reduced quality of life, underscoring the need for reliable imaging-based methods to assess muscle composition in vivo [1, 6, 7].

Chemical shift–encoded q-Dixon MRI allows accurate determination of the proton density fat fraction (PDFF), which is widely regarded as the reference standard for intramuscular fat quantification [8, 9]. CT is also attractive for muscle assessment because of its high spatial resolution, rapid acquisition, and broad availability, particularly for opportunistic evaluation in routine imaging [10, 11]. Although MRI-derived PDFF is widely regarded as the non-invasive reference standard for fat quantification, its routine clinical use remains limited by examination time, cost, availability, and contraindications in certain patient populations. In contrast, CT examinations are frequently performed in routine clinical practice and may provide opportunities for opportunistic assessment of muscle composition without requiring additional imaging. Therefore, a CT-based method capable of estimating muscle fat content could facilitate broader implementation of body composition assessment in clinical settings. However, conventional single-energy CT relies on Hounsfield unit measurements that are sensitive to scanner calibration, tube voltage, reconstruction kernel, and patient-related factors, which may limit quantitative accuracy and reproducibility [12–14]. Dual-energy CT with material decomposition has improved precision and has been applied to fat quantification in the liver, bone marrow, and skeletal muscle, but its performance remains influenced by spectral separation, noise propagation, and modeling assumptions [15–18].

Photon-counting CT (PCCT) represents a further advance in CT technology by replacing energy-integrating detector (EID) with photon-counting detector (PCD), thereby improving spatial resolution, contrast-to-noise ratio, and dose efficiency [10, 19, 20]. By recording the energy of individual photons, PCCT enables spectral sampling across multiple energy bins and the generation of high-quality spectral images, providing a technical basis for refined tissue composition analysis [20]. These properties suggest that PCCT-derived parameters may provide a stable approach for quantitative muscle fat assessment.

Accordingly, the purpose of this study was to evaluate PCCT-derived 70 keV attenuation values and virtual noncontrast fat fraction (VNC FF) for quantification of paraspinal muscle fat infiltration, with MRI PDFF as the reference standard. Specific aims were: (1) to determine agreement between PCCT-derived parameters and MRI PDFF; (2) to assess performance across different intervertebral disc levels and muscle groups; and (3) to evaluate diagnostic accuracy for fat infiltration grading and binary classification at the 30% threshold [21].

Materials and methods

Subjects

Between January and April 2025, adults with low back pain were prospectively recruited from the community and underwent both PCCT and MRI at our institution. Inclusion criteria were age ≥ 18 years, low back pain at the time of recruitment, and completion of PCCT and MRI within 2 h on the same day. Exclusion criteria were pregnancy, history of cardiac surgery, prior spinal surgery, paraspinal muscle tumours or infections, and contraindications to MRI (Fig. 1). Seventy-six participants were included (38 men, 38 women; mean age, 47.7 ± 14.0 years; mean BMI, 24.6 ± 3.3 kg/m²; Table 1). Demographic information (age, sex, height, weight) was recorded. The study was approved by the Ethics Committee of our institution (2025-037-01) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants. The order of PCCT and MRI examinations was randomized among participants.

Fig. 1.

Flowchart of participant inclusion

Table 1.

Characteristics of the included patients

Variable	Value
Number of participants (female/male)	76 (38/38)
Age (years)
Overall	47.7 ± 14.0 (25.0–80.0)
Female	49.8 ± 12.9 (26.0–71.0)
Male	45.7 ± 14.9 (25.0–80.0)
BMI (kg/m²)
Overall	24.6 ± 3.3 (19.2–34.7)
Female	23.7 ± 2.9 (19.2–31.2)
Male	25.5 ± 3.5 (19.7–34.7)

BMI: body mass index; SD: standard deviation

Continuous variables are presented as Mean ± SD (range)

PCCT protocol

All CT examinations were performed on a PCCT system (NAEOTOM Alpha VA50; Siemens Healthineers). The scan range covered the entire lumbar spine. Acquisition parameters were: tube voltage, 120 kVp with automated tube current modulation (CARE Dose4D; Siemens Healthineers); collimation, 144 × 0.4 mm; rotation time, 0.5 s; spiral pitch factor, 0.8; and slice thickness, 1 mm. The volume CT dose index was 7.4 ± 1.7 mGy, and the dose-length product was 389.7 ± 110.3 mGy·cm.

MRI protocol

MRI was performed on a 3.0-T scanner (Magnetom Vida; Siemens Healthineers) using a 32-channel spine coil. A transverse 6-echo q-Dixon sequence was acquired (TR = 9.2 ms; TE = 1.29, 2.54, 3.79, 5.04, 6.29, 7.54 ms; FOV = 340 mm; flip angle = 4°; matrix = 256 × 205; slice thickness = 3 mm; averages = 2; acquisition time = 31 s). A sagittal T2-weighted Dixon sequence was additionally acquired for qualitative spinal evaluation (TR = 3050 ms; TE = 81 ms; FOV = 260 × 260 mm; slice thickness = 4 mm; averages = 2; acquisition time = 2 min 42 s).

Image analysis

All post-processing and quantitative measurements were performed with commercial software (syngo.via, version VB60; Siemens Healthineers). For MRI, PDFF maps were automatically generated from the 6-echo q-Dixon sequence using the scanner’s reconstruction algorithm. PCCT datasets were reconstructed with a quantitative kernel (Qr40) and quantum iterative reconstruction (level 3). VNC FF was calculated using the Liver VNC module, which applies a three-material decomposition algorithm and can be used for both contrast-enhanced and unenhanced scans [22–24]. Although this algorithm was originally developed for hepatic fat quantification, it was applied in this study as a surrogate material decomposition approach for skeletal muscle, acknowledging potential tissue-specific limitations. Standardized CT attenuation values were obtained from virtual monoenergetic images (VMIs) at 70 keV. A previous phantom study demonstrated that attenuation measurements obtained from PCCT were highly consistent across different tube voltages and image quality levels, showing minimal bias at 70 keV [25].

Regions of interest (ROIs) were manually drawn at mid-disc levels from L2/3 to L5/S1, bilaterally covering the multifidus (MF), erector spinae (ES), and psoas major (PM), with reference to anatomic boundaries and along planes parallel to the corresponding intervertebral discs (Fig. 2). ROIs were initially delineated on MRI PDFF maps along the visible muscle boundaries to represent the entire muscle cross-section while excluding surrounding adipose tissue, major vessels, and imaging artifacts. Corresponding ROIs were subsequently manually drawn on PCCT images at the same anatomical level with reference to the anatomical boundaries and ROI contours identified on MRI. Quantitative parameters included MRI PDFF (reference standard), 70 keV CT values, and VNC FF.

Fig. 2.

As an example of image segmentation, the figure shows muscle segmentation with the psoas major (PM) in blue, the erector spinae (ES) in green, and the multifidus (MF) in red in MRI PDFF image (A) and 70 keV CT image (B)

Two radiologists independently reviewed all images (Reader 1, 7 years of spinal MRI experience; Reader 2, 6 years of spinal MRI experience). Prior to formal measurements, the two readers reached consensus on ROI placement criteria and anatomical boundaries. Subsequently, ROI delineation was performed independently by each reader according to the agreed protocol. All images were anonymized to ensure confidentiality. Interobserver reliability was assessed using intraclass correlation coefficients (ICCs) based on a two-way mixed-effects model with absolute agreement for average measurements. Because readers underwent protocol harmonisation before independent measurements, ICCs may be partially inflated and should be interpreted as reflecting agreement under standardised conditions. For subsequent analyses, measurements from Reader 1 were used.

Statistical analysis

Statistical analyses were performed using SPSS (version 26.0; IBM) and Python (version 3.9). Sample size estimation was performed at the participant level, assuming a correlation coefficient of 0.5 with two-tailed α = 0.05 and power = 0.90, which yielded a minimum of 37 participants. The final sample (n = 76) exceeded this requirement. Because multiple ROIs were obtained per participant and are not independent observations, we did not treat the 1824 ROIs as an independent sample size. Instead, the multilevel structure (ROIs nested within participants) was addressed using linear mixed-effects models (LMMs) with participant-specific random intercepts for regression analyses.

Normality of continuous variables was tested with the Shapiro–Wilk method. Pearson correlation coefficients were calculated for normally distributed variables; Spearman rank correlation coefficients were additionally applied for non-normally distributed data. LMMs or linear regression models were constructed with MRI PDFF as the dependent variable and either 70 keV CT values or VNC FF as predictor variables to calculate CTFF. The regression models were developed and evaluated within the same dataset and were intended to provide an internally derived quantitative relationship rather than a fully validated predictive model. Agreement between MRI PDFF and CTFF was assessed with Bland–Altman analysis. Both MRI PDFF and CTFF were categorized using identical cutoff values; agreement was assessed using Cohen’s κ, with sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy reported. Receiver operating characteristic (ROC) analysis was performed using PDFF ≥ 30% as the threshold for high fat infiltration. This threshold was selected based on a previously published muscle fat infiltration grading system, in which PDFF values ≥ 30% represent substantial fat infiltration [21].

Results

Subject characteristics

Seventy-six participants were included. Bilateral measurements were obtained at four intervertebral disc levels (L2/3–L5/S1) across three muscle groups, yielding 1824 ROIs for analysis.

Interobserver reliability

Interobserver agreement between the two radiologists was excellent, with ICCs of 0.97 for MRI PDFF, 0.96 for 70 keV CT values, and 0.95 for VNC FF.

Correlation between CT-derived parameters and MRI PDFF

ROI level analysis

Across all ROIs, 70 keV CT values showed a strong negative correlation with MRI PDFF (r = − 0.931, ρ = −0.882, all p < 0.001), while VNC FF showed a strong positive correlation (r = 0.876, ρ = 0.803, all p < 0.001). Across intervertebral disc levels, 70 keV CT values showed consistently strong negative correlations with MRI PDFF, while VNC FF demonstrated strong positive correlations: L2/3, r = − 0.853 and 0.766, ρ = −0.808 and 0.728; L3/4, r = − 0.925 and 0.865, ρ = −0.875 and 0.801; L4/5, r = − 0.946 and 0.896, ρ = −0.911 and 0.836; L5/S1, r = − 0.941 and 0.889, ρ = −0.913 and 0.813 (all p < 0.001). By muscle group, correlations were likewise strong: PM, r = − 0.825 and 0.767, ρ = −0.828 and 0.780; ES, r = − 0.954 and 0.944, ρ = −0.947 and 0.926; MF, r = − 0.964 and 0.934, ρ = −0.942 and 0.886 (all p < 0.001). Scatter plots illustrating these correlations are provided in Supplementary Figs. 1 and 2.

Individual level analysis

At the subject level, correlations were even stronger. The 70 keV CT values demonstrated consistently strong negative associations with MRI PDFF, whereas VNC FF demonstrated strong positive associations across intervertebral disc levels: L2/3, r = − 0.964 and 0.913, ρ = −0.956 and 0.900; L3/4, r = − 0.972 and 0.940, ρ = −0.956 and 0.911; L4/5, r = − 0.975 and 0.944, ρ = −0.964 and 0.921; L5/S1, r = − 0.965 and 0.952, ρ = −0.952 and 0.937 (all p < 0.001). Corresponding scatter plots are displayed in Fig. 3. By muscle group, correlations were likewise strong: PM, r = − 0.881 and 0.827, ρ = −0.881 and 0.840; ES, r = − 0.984 and 0.966, ρ = −0.978 and 0.951; MF, r = − 0.981 and 0.955, ρ = −0.972 and 0.919 (all p < 0.001), as illustrated in Fig. 4.

Fig. 3.

Scatter plots showing correlations between MRI PDFF and PCCT-derived parameters. (A–D) Correlations between MRI PDFF and 70 keV CT values at the individual level across intervertebral disc levels (L2/3–L5/S1). (E–H) Corresponding correlations between MRI PDFF and VNC FF

Fig. 4.

Scatter plots showing correlations between MRI PDFF and PCCT-derived parameters. (A–C) Correlations between MRI PDFF and 70 keV CT values at the individual level across muscle groups (PM, ES, MF). (D–F) Corresponding correlations between MRI PDFF and VNC FF

Regression models and Bland–Altman analysis

Across all ROIs, the mean MRI PDFF was 14.8 ± 7.9% (range, 1.0–58.9%). Because 70 keV CT values showed the strongest correlation with MRI PDFF, a LMM was constructed with MRI PDFF as the outcome, 70 keV CT value as the fixed effect, and subject ID as a random intercept. The resulting regression equation was: CTFF (%) = − 0.559 × 70 keV CT value (HU) + 38.5. The derived CTFF values (mean, 14.7 ± 7.3%; range, 1.5–62.0%) closely approximated MRI PDFF, as confirmed by Bland–Altman analysis (Supplementary Fig. 3).

At the individual level, regression equations stratified by intervertebral disc level were as follows: L2/3, CTFF (%) = − 0.538 × CT value (HU) + 37.1; L3/4, − 0.572 × CT value (HU) + 38.78; L4/5, − 0.602 × CT value (HU) + 40.30; L5/S1, − 0.565 × CT value (HU) + 39.53. By muscle group: PM, − 0.510 × CT value (HU) + 36.38; ES, − 0.590 × CT value (HU) + 38.87; MF, − 0.586 × CT value (HU) + 40.25. Bland–Altman analyses (Figs. 5 and 6) further demonstrated good agreement between MRI PDFF and CTFF.

Fig. 5.

Bland–Altman plots showing agreement between MRI PDFF and CTFF at the individual level across intervertebral disc levels (L2/3–L5/S1). The mean differences were − 0.007%, − 0.016%, − 0.001%, and + 0.013%, with corresponding 95% limits of agreement of − 2.39% to + 2.38%, − 2.45% to + 2.42%, − 2.73% to + 2.73%, and − 3.34% to + 3.36%, respectively

Fig. 6.

Bland–Altman plots showing agreement between MRI PDFF and CTFF at the individual level across muscle groups (PM, ES, MF). The mean differences were + 0.040%, − 0.011%, and + 0.006%, with corresponding 95% limits of agreement of − 2.71% to + 2.79%, − 2.50% to + 2.48%, and − 2.80% to + 2.82%, respectively

Muscle fat infiltration grading and diagnostic performance

Using thresholds of < 10%, 10–30%, 30–50%, and > 50%, MRI PDFF and CTFF were classified into four grades of fat infiltration [21]. Overall agreement was 82.0% (1495 of 1824), with a Cohen’s κ of 0.623. Most discrepancies occurred at the boundary between grade 1 and grade 2. Sensitivity was highest for grade 2 (0.883), indicating that CTFF was most effective for detecting 10–30% infiltration. For grades 3 and 4, specificity and NPV exceeded 0.98, demonstrating high reliability for ruling out 30–50% and > 50% infiltration. However, the sample size in the > 50% group was limited, which may have reduced stability. Detailed results are provided in Supplementary Tables 1 and 2. For binary classification using a 30% threshold, CTFF showed excellent diagnostic performance for identifying MRI PDFF ≥ 30%, with an AUC of 0.993 (95% CI: 0.990–0.997) on ROC analysis (Supplementary Fig. 4).

Discussion

This study demonstrates that PCCT-derived 70 keV CT values are strongly associated with MRI PDFF for quantifying paraspinal muscle fat infiltration across disc levels and muscle groups. The 70 keV CT values demonstrated strong and consistent associations with MRI PDFF across disc levels and muscle groups, whereas VNC FF showed greater variability. These findings suggest that spectral information derived from PCCT may provide a quantitative surrogate for MRI-based fat assessment within the studied cohort.

Conventional single-energy CT primarily relies on HU values as indirect surrogates of muscle density. However, HU values are easily influenced by acquisition parameters, body habitus, and scanner variability, thereby limiting quantitative accuracy and reproducibility. By contrast, VMI enables acquisition of CT values at a defined energy level that are theoretically independent of scanner type and tube voltage [26]. This advantage derives from the count-weighting properties of PCD and the universality of energy-dependent attenuation coefficients [26, 27]. Compared with EID-based DECT, PCCT provides superior noise performance and improved noise texture, spatial resolution, and detectability across all low-keV VMIs [28].

In the present study, although VNC FF correlated well with MRI PDFF, its accuracy was more susceptible to the inherent limitations of material decomposition modeling, including dependency on modeling parameters, spectral overlap, and noise propagation [29]. Furthermore, the spectral attenuation profiles of water and fat largely overlap, which inherently limits their precise separation using x-ray energy alone [30].

From an anatomic perspective, PCCT-derived parameters correlated most strongly with MRI PDFF in the ES and MF, while correlations were slightly lower in the PM. The PM consists of multiple overlapping segmental fascicles, each originating from the intervertebral discs or transverse processes, with relatively loose fascicular arrangement near the lumbar spine, which may increase ROI delineation error [31]. In contrast, the ES and MF have clearer anatomic boundaries, facilitating consistent ROI placement. Furthermore, the relatively small number of subjects with severe fat infiltration (> 50%) resulted in a skewed distribution, which may have influenced regression slope and intercept.

For categorical assessment, PCCT-derived CTFF demonstrated good agreement with MRI PDFF (κ = 0.623). Particularly in binary classification using the 30% threshold, CTFF achieved excellent diagnostic accuracy, with an AUC of 0.993. However, agreement across all fat infiltration categories was more modest, and performance in the > 50% category was limited by the small number of cases. These findings suggest that CTFF may currently be more suitable for classification using the 30% threshold than for differentiating among all fat infiltration grades. These findings indicate that PCCT-derived quantitative parameters can be used to assess muscle fat content in routinely acquired CT data. The present study was primarily intended to validate the technical relationship between PCCT-derived parameters and MRI PDFF rather than to evaluate clinical or prognostic significance. However, the clinical implications of such opportunistic measurements remain to be established and should be evaluated in studies incorporating functional outcomes, longitudinal follow-up, or prognostic endpoints.

Although PCCT is not yet widely available in routine clinical practice and involves exposure to ionizing radiation, it provides several advantages over conventional CT, including higher spatial resolution, lower electronic noise, and improved dose efficiency [19, 32]. Compared with MRI, PCCT abdominal imaging covers a larger field of view, allowing simultaneous quantification of visceral and subcutaneous fat as well as abdominal wall musculature—parameters closely linked to clinical outcomes. While its current role in opportunistic screening remains limited, our findings provide preliminary evidence supporting the feasibility of simple and accurate muscle fat quantification based on PCCT parameters. With broader clinical adoption of PCCT, its use for opportunistic muscle fat quantification may gradually increase.

A methodological strength of this study is the use of linear mixed-effects models, which accounted for the clustered structure of ROI measurements and avoided treating multiple ROIs from the same participant as independent observations. This study has several limitations. First, measurements were based on manually delineated ROIs, which may introduce observer-dependent variability despite high interobserver agreement; test–retest reproducibility was not assessed, and repeatability across repeated scans or acquisition conditions warrants further investigation. Second, the regression equation used to derive CTFF was developed and evaluated within the same dataset, and the number of participants with severe fat infiltration (> 50%) was relatively small; external validation in independent cohorts is therefore required to confirm robustness and statistical stability. Finally, generalisability may be influenced by scanner-related factors. Although 70 keV virtual monoenergetic images are theoretically less dependent on tube voltage, attenuation values and derived regression equations may still be influenced by reconstruction kernels, iterative reconstruction settings, and differences among PCCT platforms or future scanner generations. Because the proposed CTFF equations were derived using a single vendor PCCT system with a specific reconstruction kernel (Qr40) and acquisition protocol, their generalisability remains uncertain. Further multicentre studies across different platforms and reconstruction settings are required before broader application. The proposed CTFF equations are intended to describe the relationship between PCCT-derived attenuation values and MRI PDFF within the present cohort and should not be considered universally applicable without external validation.

Conclusions

This study demonstrates the methodological feasibility of PCCT-based quantification of paraspinal muscle fat, with PCCT-derived 70 keV attenuation values showing strong agreement with MRI PDFF and supporting internally derived regression–based estimation of muscle fat fraction. In addition, the internally derived CTFF exhibited good diagnostic performance for grading fat infiltration. Together, these findings support the technical feasibility of PCCT for quantitative muscle fat assessment. The clinical and prognostic relevance of these measurements remains to be established and should be investigated in future studies incorporating functional and longitudinal outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

CTFF

CT fat fraction

EID

Energy-integrating detector

ES

Erector spinae

ICCs

Intraclass correlation coefficients

LMMs

Linear mixed-effects models

MF

Multifidus

PCCT

Photon-counting CT

PCD

Photon-counting detector

PDFF

Proton density fat fraction

PM

Psoas major

ROC

Receiver operating characteristic

ROIs

Regions of interest

VNC FF

Virtual noncontrast fat fraction

VMIs

Virtual monoenergetic images

Author contributions

The study was conceived by DBS, designed by ZHZ and DBS. DBS drafted the first version of the manuscript. JPW and DL offered advice on the workflow and critically revised the manuscript. JPW and LZ provided insightful comments and revised the manuscript. All authors have read and agreed to the published version of the manuscript.

Funding

This study has received funding by the Key Laboratory of Functional Molecular Imaging of Tumor and Interventional Diagnosis and Treatment of Shaoxing City, and the Zhejiang Medical and Health Technology Project (2024KY1703).

Data availability

With permission from Shaoxing people’s Hospital upon reasonable request, the dataset is available from the corresponding authors Dr. Zhenhua Zhao. Restrictions do exist, the information isn’t available to the general public.

Declarations

Ethics approval and consent to participate

This study was approved by the research ethics committee of Shaoxing people’s Hospital (No. 2025-037-01) and was in accordance with the Declaration of Helsinki. Written informed consent was obtained from all subjects (patients) in this study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.