This retrospective in-silico study was conducted in accordance to the guidelines of the local institutional review board. Anonymised CT datasets of one participant with a healthy liver (n = 1) and four patients, each of them with characteristic liver lesions (HCC, haemangioma, cyst, and metastasis) were used as a template for the numerical experiment. The diagnosis of liver lesions was confirmed by biopsy for HCC and metastasis while follow-up studies including magnetic resonance imaging confirmed the diagnosis of haemangioma and cyst.
Dual-contrast injection protocol
For dual-contrast liver imaging, we defined a dedicated injection protocol (Fig. 1) with a sequential application of two different CAs. The attenuation curve of CA1 (gadolinium-based CA) within a region of interest (ROI) in the abdominal aorta was used as a test bolus to determine the individual patient-specific timing of the blood circulation and the maximum of arterial contrast enhancement. This information is essential to precisely forecast the time point of maximal arterial enhancement of the liver by CA2 (iodine-based CA).
At time point T0, CA1 (Magnograf®, Bayer Pharma AG, Berlin, Germany, administered at a dose of 0.2 mL/kg body weight, with an injection rate of 4 mL/s) is injected generating a maximal arterial enhancement of the liver at time point T1, followed by injection of CA2 at T2 (Iomeron 370, administered at a dose of 1 mL/kg body weight, with an injection rate of 4 mL/s). The time difference ΔT = T1 – T0 represents the period from CA injection to maximum of arterial enhancement. Sequential injection of CA1 and CA2 leads to a dual-phase dual-contrast distribution at the time point T3 = T2 + ΔT, with CA1 in the portal venous and CA2 in the arterial phase (see Fig. 1). At time point T3, the SPCCT examination is performed to simultaneously assess the contrast distribution of both CAs in the liver at different phases.
Numerical SPCCT experiment
The numerical experiment was set up to be as realistic as possible by modelling both the energy-dependent attenuation characteristics of the patients and the actual physical performance of a SPCCT scanner. The number of detected photons in a detector pixel mainly depend on the three major components: the x-ray tube spectrum, the detector response and the attenuation properties of the patient [7]. In order to mimic an SPCCT acquisition, a realistic x-ray tube spectrum model [8] and a realistic detector response model (measurements at the European Synchrotron Radiation Facility [ESRF] in 2014, not published) were utilised. The resulting photon counts for each x-ray were calculated using the approach presented by Roessl and Proksa [9]. Poisson-distributed noise was added to the calculated photon counts. The simulation parameters match an already existing SPCCT prototype, where axial scans over 360° are obtained with a tube current of 50 mA, a tube voltage of 120 kVp, scanner rotation time of 1 s and 2400 projections per rotation. The noise threshold in the detector modelling was set to 30 keV and all images were reconstructed on a voxel grid of 0.39 × 0.39 × 0.25 mm [10].
Preparation of numerical liver phantom and liver lesions
To gain information of the attenuation characteristics in a real patient, we used an axial slice of a multi-phase liver CT scan and substituted the liver with a synthetic liver model. For this, we segmented the CT image by thresholding the data into a bone and soft tissue image (Fig. 2), followed by manual segmentation and replacement with a synthetic liver model with homogeneous CT values of 50 HU. A total of four common liver lesions have been selected from the clinic’s patient database by an experienced radiologist (D.M. nine years of experience).
Characteristics of arterial and portal venous contrast enhancement of these lesions were considered as follows (Fig. 3). The cyst does not enhance in both the arterial and the portal venous phase with a CT value of about 0 HU. HCC typically shows an arterial hyper-perfusion in the arterial contrast phase (+30 HU compared to liver enhancement) and a washout phenomenon in the venous phase (– 15 HU compared to liver enhancement). Haemangioma typically presents with a ‘closing iris’ pattern of enhancement, with a peripheral nodular hyper-vascularisation in the arterial phase (+75 HU compared to liver enhancement) and a centripetal contrast filling in the venous phases (+55 HU compared to liver enhancement). The metastasis (typically from colorectal cancer) shows a peripheral rim-like enhancement in the arterial phase (100 HU), which is further increased in the portal venous phase images (120 HU). The core of the metastasis was simulated without enhancing in arterial phase images and a little uptake in portal venous phase (35 HU). The liver lesions were added into the homogenous synthetic liver model, each with three different sizes: 5 mm, 10 mm and 20 mm, respectively. The 20 mm and 10 mm liver lesions were positioned in the right lobe of the liver in segment VIII and segment VII and the 5 mm lesions had their position in segment II according to Couinaud’s system of liver anatomy [11]. The enhancement of healthy liver was 60 HU in the arterial and 90 HU in the portal venous phase.
Material decomposition
We employed a projection-based maximum-likelihood method for the material decomposition into water, CA1, and CA2 [7]. This method uses the photon x-ray spectrum, the linear attenuation coefficients of water, CA1, and CA2 together with the spectral detector response function [7, 9]. The decomposition algorithm estimates the material composition that fits best to the simulated noisy photon counts for each projection and detector pixel. More precisely, the multi-bin data were pre-processed and an integrated conventional CT image was obtained. After further corrections, a maximum likelihood material decomposition of the attenuation into a water, iodine, and gadolinium material basis was performed (see below).
To test the feasibility of spectral CT liver imaging with respect to its clinical applicability, we carefully adjusted the dose to be comparable to that used in conventional CT liver imaging protocols. Therefore, a conventional CT simulation using an energy-integrating detector of the chosen dataset was carried out to determine the nominal x-ray dose yielding an image noise of approximately 20 HU in the liver using the same reconstruction parameters as for the SPCCT simulation [10].
Image reconstruction and processing
The outcome of the material decomposition are three material projection datasets (CA1, CA2, and water). The anti-correlated noise in the material images is suppressed by an iterative image based statistical de-noising algorithm [11]. Pixel-by-pixel based image analyses (e.g. cluster analysis or support vector machine) can be applied to the final material images to make best use of the simultaneously acquired arterial and portal venous phases. The three materials CA1, CA2, and water can be understood as a three-dimensional (3D) vector space. Each point \( \overrightarrow{x} \) in this space represents a different combination of water, CA1, and CA2. Image pixels belonging to the same tissue type form clusters in the 3D vector space. In case the clusters do not or only partially overlap, the different types of tissue can be identified. We model a cluster for tissue type t (e.g. HCC, metastasis…) by a 3D joint real normal distribution, as follows:
$$ {p}^t\left(\overrightarrow{x}\right)={\prod}_{i=\left\{ water, CA1, CA2\right\}}\frac{1}{\sqrt{2\pi }{\sigma}_i^t}\mathit{\exp}\left(-\frac{1}{2}{\left(\frac{x_i-{\mu}_i^t}{\sigma_i^t}\right)}^2\right), $$
where: x
i
with i = {water, CA1, CA2} are the entries of \( \overrightarrow{x} \) describing a point in the 3D vector space; \( {\mu}_i^t \) with i = {water, CA1, CA2} are the coordinates of the centre of the cluster for tissue type t; and \( {\sigma}_i^t \) with i = {water, CA1, CA2} are the standard deviations of the cluster along the material axes.
Correlated noise between the material images was suppressed to a level allowing to neglect it in \( {p}^t\left(\overrightarrow{x}\right) \). The value \( {p}^t\left(\overrightarrow{x}\right) \) describes the likelihood for a point \( \overrightarrow{x} \) to belong to the tissue type t. Thus, displaying \( {p}^t\left(\overrightarrow{x}\right) \) for each image pixel is called likelihood map for tissue type t. The scatter plot and the likelihood representation of the decomposed material data were used to proof the lesion detection capability.