RPS 513 - Unlocking the power of photon-counting CT

Author Block: J. Wucherpfennig¹, L. T. Rotkopf¹, H-P. Schlemmer¹, M. Frölich², S. O. Schönberg², S. Sawall¹, M. Kachelrieß¹; ¹Heidelberg/DE, ²Mannheim/DE
Purpose: To evaluate the separate and combined influence of the small pixel effect (SPE) and tin prefiltration in terms of signal-to-noise ratio at unit dose (SNRD) in energy-integrating CT (EICT) and photon-counting CT (PCCT).
Methods or Background: SPE refers to higher SNRD for scans with small detector pixels versus larger pixels at the same MTF. Tin effect refers to the selective removal of low-energy photons before reaching the patient. Fully exploiting the tin effect requires adjusting tube voltages, which was not possible in our setting with the same tube voltages and spectra on both scanners. Both effects improve the SNRD. To quantify, images of semi-anthropomorphic abdomen phantoms of sizes 20×30 cm (S), 25×35 cm (M), 30×40 cm (L), and 35×45 cm (XL) were acquired with an EICT (Somatom Definition Flash, Siemens Healthineers) and a PCCT (Naeotom Alpha.Peak, Siemens Healthineers) at 140 kV and 140 kV Sn. PCCT scans were done in standard and ultra-high-resolution mode to exploit the SPE. All reconstructions within each scanner were performed with the same MTF.
Results or Findings: SPE allowed for 36% (S), 30% (M) and 6% (L) dose reduction. Tin effect in PCCT was observed only in the XL phantom, with 25% dose reduction. The unobserved tin effect in smaller phantoms and the reduced SPE in larger phantoms are presumably due to a proprietary raw-data-dependent filter implemented in the PCCT. Tin effect in EICT was consistent across all phantom sizes with a dose reduction of 20% (S), 29% (M) and 30% (L).
Conclusion: EICT benefits from the tin effect. PCCT allows to exploit the small pixel effect. Making use of the expected combined effect in PCCT, however, seems to require an adaptation of the vendor-proprietory rawdata filter.
Limitations: Findings are based on phantoms.
Funding for this study: None.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: X. Liu, F. Dong, K. Li; Houston/US
Purpose: In the NAEOTOM Alpha photon-counting detector CT (PCD-CT), the mean CT number of a given tissue has shown good consistency across varying dose levels. This observation has been interpreted as evidence that pulse pileups are negligible in Alpha. However, such stability in CT numbers can also result from effective pileup correction. The aim of this study was to evaluate the impact of pileup correction on the signal and noise characteristics of the Alpha PCD-CT.
Methods or Background: Images were acquired using an Alpha scanner at four tube potentials and twelve tube current settings. The vendor's pulse pileup correction was automatically applied during reconstruction. Experimental measurements included image variance, noise power spectrum (NPS), MTF, and mean HU. The dependence of each metric on tube current and dose level was analyzed.
Results or Findings: The MTF, the shape of the NPS, and the mean HU showed no significant variation with tube current or dose. However, the image variance deviated notably from the classical inverse relationship with tube current (variance ∝ 1/mA). Specifically, the observed variance was higher than predicted by this relationship, with the deviation becoming more pronounced at higher mA levels. At tube currents above 300 mA at 120 kV, corresponding to a flux of 1E8 counts-per-second/pixel, the variance even increased with mA. This counterintuitive behavior can only be explained by the presence of a pileup correction mechanism.
Conclusion: Pulse pileup correction was automatically applied in the Alpha PCD-CT imaging chain. This correction linearizes the mean detector counts with respect to tube current but results in elevated image noise. The correction appears to be applied on a pixel-by-pixel basis, as it does not affect spatial resolution or alter the noise texture.
Limitations: The study is based only on phantoms.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: E. Pimenta¹, G. Boiset¹, G. Negro¹, M. Bettencourt¹, A. Tomal², L. J. Oostveen³, M. Caballo⁴, I. Sechopoulos³, P. R. Costa¹; ¹São Paulo/BR, ²Campinas/BR, ³Nijmegen/NL, ⁴'s-Hertogenbosch/NL
Purpose: To evaluate the effect of noise texture at different dose levels on volumetric accuracy and precision of synthetic solid nodules (SNs) and ground-glass opacities (GGOs) imaged with photon counting CT (PCCT).
Methods or Background: A 3D-printed SN and a GGO were inserted into an anthropomorphic lung phantom. Ground-truth volumes were quantified using a µCT (U-SPECT6CThr, MILabs). The phantom was scanned five times with repositioning in a prototype PCCT system (Canon Medical Systems) at four CTDIvol levels (0.3, 0.5, 0.8 and 1.1 mGy). Images were reconstructed using hybrid iterative reconstruction (AIDR 3D/FC54 kernel).
Volumes were segmented (3DSlicer, v5.6.2) by readers blinded to the imaging parameters. Accuracy was assessed using median relative error and Bland-Altman plots. Precision was evaluated using the coefficient of variation (CV, IQR-based). Noise power spectra (NPS) were calculated from four ROIs around each nodule. Associations between volumetric performance, NPS features (area, downslope σ, peak frequency fpeak, and average frequency fave) and CTDIvol were evaluated using ordinary least squares and quantile regression (QReg, τ=0.5).
Results or Findings: SN volumetry showed strong dose-dependence for accuracy (QReg R²=0.91; CTDIvol β=0.93, p=0.034), whereas GGO accuracy modeling was weaker (R²=0.46). In contrast, GGO precision was dose-dependent (R²=0.57; β=7.71, p=0.022). For SN, precision strongly correlated with NPS features (QReg R²=0.78), including fpeak (β=−154.8, p<0.001), fave (β=170.5, p<0.001), and σ (β=−151.6, p<0.001).
Conclusion: For the SN, accuracy was dose-driven while precision was associated with noise texture. For the GGO, precision was predominantly dose-dependent. These findings highlight the importance of tailored protocol optimization in PCCT.
Limitations: A limited number of nodules and a single reconstruction method/kernel combination, which restrict generalizability.
Funding for this study: The authors would like to acknowledge financial support of Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP) (grants 2018/059820, 2021/14688-0, 2022/114570 and 2023/03945-8), Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) (grants 141335/2021-1, 131691/2021-0, 138533/2022-9, 302986/2023-5 and 311657/20214).
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: K. Torfs, K. Merken, D. Petrov, H. Bosmans; Leuven/BE
Purpose: Mesh-based resolution algorithms quantify sharpness on patient images via modulation transfer functions (MTF) taken along the patient air-skin interface, but often fail for noisy chest images (e.g. sharp kernels/low dose). We introduce a high-noise-adapted mesh algorithm (AM), with altered edge-profile conditioning, and compare it to the current mesh algorithm (CM) and a phantom-based circular-edge method (CE), considered gold standard.
Methods or Background: High-contrast circular phantom images were simulated with resolution and noise texture characteristics derived from analytical ground-truth MTFs. Ideal coefficients for the analytical model were fitted to the CE MTF of nine Siemens SOMATOM Edge kernels.
Ten repeated stacks were simulated per combination of:
• Kernel (Br32–Br60, Qr32–Qr60, Bl57)
• Simulated phantom tilting (0°-10° in 2° increments)
• Noise magnitude (0-200HU in 25-HU steps)
• Clothing (present/absent)
Percentage error (PE) was compared across CE, AM and CM.
Results or Findings: CE performance was largely noise-independent, but sensitive to high tilting (>6°: PE=60±40%). For tilting <6° (PE=2.5±1.9%), CE was more accurate (p<0.001) than both AM/CM, while AM outperformed (p<0.001) for tilting >6°.
Mesh-based algorithms were tilt-resistant but noise-sensitive:
• Low noise (0-50 HU): PE of AM lower than CM (p<0.001: 3±3% vs. 5±3%)
• Medium noise (75-125 HU): PE of AM lower than CM (p<0.001: 8±5% vs. 40±30%)
• High noise (150-200 HU): PE of AM lower than CM (p<0.001: 17±9% vs. 70±30%)
Similar tendencies were observed with clothing present, only at slightly higher PE, and CM outperformed AM (p<0.05) for noises 0-25 HU.
Conclusion: For patient-based resolution measurements, tested on non-circular objects, AM reduces error significantly versus CM in most scenarios, limiting PE to maximally 10% for noise up to 125 HU. CE remains optimal for circular phantoms, used in quality control, given precise positioning.
Limitations: None
Funding for this study: Funding was provided by Kom op Tegen Kanker (G0B1922N) in light of a doctoral grant
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Approved under internal reference number S68527

Author Block: G. Kristof, R. Von Stritzky, E. Eulig, M. Kachelrieß; Heidelberg/DE
Purpose: To improve AI-based CT image denoising, aiming at more naturally looking output images with less hallucinations, by providing noise covariance information to denoising neural networks.
Methods or Background: CT image noise is correlated between neighboring voxels. The correlation strongly depends on the patient’s global attenuation properties. We provide this information to deep denoising neural networks by simulating several additional noise-only images as additional network input. Up to ten of these noise only images and the low-dose image, which is the one to be denoised, are now input into a denoising network. We tested this noise-augmented deep denoising (NADD) for the CNN10, ResNet, and WGAN-VGG denoising networks. Our training, validation and test dataset consists of approximately 50,000 clinical CT images. For each network, we compared the NADD version with its original version.
Results or Findings: The NADD versions of the networks performed significantly better than the original versions. The structural similarity index measure (SSIM) improved from 0.937 to 0.956 for CNN10, from 0.942 to 0.954 for ResNet and from 0.917 to 0.940 for WGAN-VGG. In particular, a significant qualitative improvement of the images is observed and far less hallucinations are apparent with NADD.
Conclusion: Using additional noise-only realizations as network input significantly improves the denoising capability of denoising networks. A downside is the requirement of additional image reconstructions corresponding to the number of additional noise realizations used. NADD may help to lower the effective dose to and therefore cancer risk of the patient.
Limitations: Our study only considers 2D, image-by-image, denoising. Future work should extend the denoising networks to true 3D networks that can denoise a whole volume in a 3D fashion.
Funding for this study: This study was supported in parts by the German Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV) under grant 67KI2036B, and by the Helmholtz International Graduate School for Cancer Research, Heidelberg, Germany.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: A. Kazhybekova, J. Cooke, M. Kelly, I. Hernandez-Giron; Dublin/IE
Purpose: To present an open-access CT image quality evaluation platform for automated comparison of reconstruction algorithms (AI/DLR vs iterative), based on structural similarity index (SSIM) and histogram analyses.
Methods or Background: CT-DImQ_SSIM is a Python-based GUI (PySide6, scikit-image library) that evaluates SSIM (Wang2004) between two images/stacks, with user-selected SSIM parameters and input window level/width mimicking different visualization conditions. SSIM local maps are automatically created globally and in selected VOIs and saved (TIFF), and average SSIM-values calculated. As an application the Kyoto Kagaku lung paediatric phantom with 3D-printed nodules was scanned (GE-Revolution, standard-thorax, clinical iterative reconstruction, ASIR50-lung and DLR, TrueFidelityHigh). SSIM analysis was performed in selected VOIs (nodules, vessels, rib, vertebra, heart).
Results or Findings: High similarity between both reconstructions (82% pixels, 0.8Conclusion: CT-DImQ_SSIM is an open-access tool for fast and automated SSIM CT image analysis, for imaging protocol and reconstruction algorithms comparison. SSIM maps are a useful visual and quantitative tool to identify potential areas in the image where CT reconstructions/settings may differ, with potential use to identify DLR hallucinations or significant CT-value changes. In an anthropomorphic thorax phantom study, DLR preserved image structural information overall, whereas in edges, especially in high contrast regions it showed low similarity compared to the standard iterative reconstruction.
Limitations: The phantom lacks lung parenchyma which can introduce a bias in comparison with patients reconstructed with same algorithms.
Funding for this study: Ainur Kazhybekova holds an UCD Ad astra PhD Scholarship.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: R. Coulibaly¹, A. Robert¹, E. Thomas¹, C. A. Hernandez-Fajardo¹, A. J. Gutwinska¹, M. N. Antonuccio², P. Ravipati¹, J. Greffier³, S. A. Si-Mohamed¹; ¹Villeurbanne/FR, ²Suresnes/FR, ³Nîmes/FR
Purpose: To assess the impact of ultra-high resolution (Detailed2) and high-resolution (HRB) reconstruction kernels on K-edge imaging of ytterbium and gadolinium using spectral photon-counting CT (SPCCT).
Methods or Background: A custom-made cylindrical phantom containing five fillable holes was used. Two configurations were tested: configuration1 contained holes with Gd (0.5, 1, 2, and 10mg/mL) and configuration2 contained holes with Yb at the same concentrations. For each configuration, nine helical acquisitions were performed on the phantom at 120kVp and 75mAs. Conventional images, as well as color K-edge images of Yb (61keV) and Gd (50.2keV), were reconstructed using the HRB and Detailed2 kernels. A three-basis (photoelectric/Compton/K-edge) projection-based material decomposition was used to obtain the K-edge images. Image quality was evaluated in terms of noise power spectrum (NPS), and task-based transfer function (TTF).
Results or Findings: In conventional images, the Detailed2 kernel resulted in higher noise (34.87UH vs 61.57UH configuration1; 34.90UH vs 61.55UH configuration2) and shifted the NPS peak to higher spatial frequencies (0.30mm⁻¹ vs 0.80mm⁻¹ configuration1; 0.33mm⁻¹vs 0.80mm⁻¹ configuration2) compared with HRB.

In K-edge images, both kernels showed similar noise (0.07mg/mL gadolinium; 0.08mg/mL ytterbium) with stable fₚₑₐₖ values (0.33mm⁻¹ gadolinium; 0.35 vs 0.32mm⁻¹ ytterbium).

In conventional images, Detailed2 provided higher spatial resolution. As concentration increased from 0.5 to 10mg/mL in configuration2, f₅₀ ranged 0.242–0.586mm⁻¹ with Detailed2 and 0.247–0.459mm⁻¹ with HRB. In configuration1, Detailed2 ranged 0.330–0.545mm⁻¹, HRB 0.295–0.448mm⁻¹.

In K-edge images, HRB performed better at low concentrations. At 0.5mg/mL, f₅₀ was 0.239 vs 0.231mm⁻¹ for ytterbium and 0.280 vs 0.250mm⁻¹ for gadolinium. At 10mg/mL, Detailed2 outperformed HRB: for ytterbium, 0.732 vs 0.535mm⁻¹; for gadolinium, 0.626 vs 0.451mm⁻¹.
Conclusion: For K-edge imaging, noise was similar, but Detailed2 gave higher resolution at high concentrations, while HRB was slightly better at low concentrations.
Limitations: Phantom study.
Funding for this study: This work was supported by the ERC starting Grant KOLOR SPCCT Imaging (N°101118079).
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: A. Tognella; Zürich/CH
Purpose: Liver fat quantification on CT is often confounded by hepatic iron deposition and the use of iodinated contrast agents. This phantom study aimed to assess the feasibility and accuracy of quantifying liver fat content (LFC) in the presence of iron using pre-contrast spectral localizer radiographs (SLR) from a photon-counting detector CT (PCD-CT).
Methods or Background: Sixteen phantoms were created using mixtures of liver tissue, fat, and iron to simulate four levels of LFC (0%, 10%, 30%, 50%) and four levels of liver iron concentration (LIC: 0, 1.5, 3, and 6 mg/mL). All phantoms were scanned on a PCD-CT using three tube current settings (10 mA, 50 mA, 300 mA), acquiring SLRs. Material decomposition of high- and low-energy bin data yielded water and hydroxyapatite (HA) maps. HA-values were analyzed as a function of LFC and LIC, and water-values were correlated with corresponding HA-values.
Results or Findings: Increasing LFC resulted in a linear decrease in HA-values, consistent across LIC levels (slopes ranging from -0.0016 to -0.0023, r = 0.997 – 1.0). Conversely, increasing LIC led to a linear increase in HA values, independent of LFC (slopes ranging from 0.0147 to 0.017, r = 0.978 – 1.0). When combined with water values in a two-dimensional material space, these consistent linear trends enabled accurate estimation of LFC regardless of LIC. Findings were reproducible across all tube current settings.
Conclusion: This phantom study demonstrates that SLRs from PCD-CT enable accurate estimation of liver fat content, even in the presence of iron deposition. If validated in vivo, this technique may facilitate opportunistic screening for hepatic steatosis and iron overload in clinical practice solely using SLRs.
Limitations: Proof-of-principle study with phantom setup offers simplified conditions. No evaluation of patient data. Limited to a specific PCD-CT.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: L. G. Pedersen¹, A. Schulz¹, A. Dybwad¹, D. Crotty², H. Linder³, J. Levy⁴, K. Jensen¹; ¹Oslo/NO, ²Cork/IE, ³Stockholm/SE, ⁴Greenwich, NY/US
Purpose: Smaller metastatic liver lesions can be difficult to detect on conventional energy-integrating (EID) CT. The improved spatial and spectral resolution of Photon Counting CT (PCCT) is expected to improve iodine contrast detection in liver imaging.
Methods or Background: A semi-anthropomorphic liver phantom (PhantomLab) was scanned using GSI on a Revolution Apex EID CT and subsequently on a 4th-generation prototype silicon-based PCCT system (both GE HealthCare). The phantom contains both homogeneous inserts without Iodine and Iodine inserts of varying concentration (1-3mg/mL) to mimic metastatic liver lesions of diameter 3-15mm. Similar CTDIvol levels (5-25mGy) were used to image on both systems with other technique parameters prescribed as similar as possible.
Using prototype image reconstruction algorithms, Virtual Monoenergetic Images (VMI) from 40-70keV and Iodine and Water material maps were reconstructed. Voxel-based Hounsfield Unit (HU), image noise (measured as standard deviation (SD)), and Contrast-to-Noise (CNR) were quantified. A two-alternative forced choice reader study was performed to visually detect iodine inserts.
Results or Findings: For the 15mm lesion, PCCT VMI reconstructed using the prototype image reconstruction process demonstrate increased CNR as CTDIvol increases and keV reduces. At 70keV, the CNR of the 15mm lesion in PCCT increased from 1.5-3.3 with increasing CTDIvol. At equal CTDIvol (10mGy), PCCT images (8.9-12 HU) showed a significant decrease in background image noise range relative to EID images (11.6-35.1 HU). While PCCT image noise is reduced, the iodine HU stays consistent across CTDIvol, resulting in up to 60% increased CNR for all reconstructed VMIs on the PCCT relative to EID.
Conclusion: Reader study participants observed the 3 mm lesion in 90% of cases in PCCT images obtained between 15 and 20 mGy.
Limitations: Only phantom study, results should be confirmed in a patient study.
Funding for this study: None.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: