Feasibility of unconstrained three-material decomposition: imaging an excised human heart using a prototype silicon photon-counting CT detector

It is well known that the development efforts of photon-counting detectors (PCDs) for computed tomography (CT) are motivated by a desire to resolve some of the fundamental problems with single-energy CT using energy-integrating detectors. These include limited spatial resolution, limited low-dose performance [1], and non-quantitative imaging [2]. The current proposition of addressing the issue of non-quantitative imaging clinically is dual-energy CT (DECT). The use of DECT can provide quantitative tissue-specific information as a decomposition into two basis materials [3].

It is also well known that the linear attenuation coefficient of human tissue is well characterized by two basis materials and that a third basis function can be used to capture the linear attenuation of heavier elements with k-edges in the spectrum. Since dual-energy systems only provide measurements at two effective energies, some kind of constraint (typically mass preservation) needs to be imposed in order to estimate a third component. Although this enables visualization of k-edge materials with dual-energy systems, any deviation in the imaged object from the assumptions of the constraint results in bias in the decomposition [4], which in practice can result in poor differentiation of iodine and calcium.

Iodine, used in the majority of CT scans, has the k-edge at 33 keV which is generally considered too low to allow k-edge imaging due to the low fraction of x-rays below this energy after attenuation through a patient [2, 5, 6]. Instead, contrast agents such as gadolinium and gold have been proposed. Iodine content can be determined with PCDs by using constraints, but like for dual-energy methods, this can lead to bias.

The demand for and general trend towards non-invasive coronary plaque diagnostics are evident in the introduction of, e.g., coronary artery calcium score CT as opposed to angiography. There is a large patient population that presents with acute chest pain, normal initial biochemical markers, and normal or non-diagnostic electrocardiogram [7, 8] that could potentially benefit from improved coronary plaque diagnostics since absence of stenosis and/or primarily calcified plaque has been shown to be a good predictor of low-risk individuals [9, 10]. However, the atherosclerotic disease is a heterogeneous process with different stages and there is an inter-individual difference in risk and possibly even intra-individual risk [11]. Therefore, increasing imaging efforts are made to find patients with high-risk/vulnerable plaques [12, 13]. One such approach is to try and better differentiate calcium and iodine using spectral photon-counting CT.

There are a number of PCD systems currently being evaluated on phantoms and human volunteers, based on either cadmium telluride [6, 14‐18] or silicon [19‐21], and spectral photon-counting CT is expected to be clinically available within the next 5–10 years. One of the challenges driving the design of these detectors is how to handle pulse pileup [22] at high flux. Silicon detectors have a technical advantage that allows for handling high count rates without degrading the energy resolution [23, 24]. It has been shown in simulation that silicon detectors can, depending on the system design, perform better than cadmium-based detectors on spectral imaging tasks such as iodine quantification [25].

We want to challenge the view that iodine cannot be more accurately decomposed in the presence of calcium with PCDs than with dual-energy methods ([6], p. 32), by proposing three-material decomposition as an application for imaging tasks where better separation of calcium and iodine is feasible and would add clinical value. We have in previous publications shown that decomposition into more than two basis materials is feasible in phantom studies with this detector [20, 21]. This has also been demonstrated with other PCDs with phantoms [26‐31], anatomical specimens [27, 28, 32], and animal models [33, 34]. The purpose of this work is to provide proof of concept that unbiased three-material decomposition with iodine can be achieved in a clinical specimen with human tissue with an experimental photon-counting silicon detector in a bench-top setup. To this end, we image an ex vivo excised heart with calcified plaque to evaluate the potential for added clinical value for imaging tasks where differentiation of iodine and calcium is of clinical relevance.

The PCD images have noticeably higher spatial resolution than the images from the clinical scanner. This is explained both by the smaller pixel size of the PCD (0.270 mm compared to 0.625 mm in the isocenter) as well as the smaller focal spot of the bench-top setup (0.4 mm compared to 1.2 mm). The contribution from each of them to the resolution improvement can be estimated by approximating the combined blur from the focal spot size and detector pixel size as \( B\left(F,D\right)=F\left(1-\frac{1}{M}\right)+\frac{D}{M} \), where F is the focal spot size, D is the detector pixel size, and M is the magnification, inserting values yields that

The total reduction in blur is thus approximately 61%, of which 30% is due to the reduced pixel size and 31% is due to the reduced focal spot size. The PCD images are also noisier than the images from the clinical scanner, which is expected due to its higher spatial resolution and the reduced dose. It is well known that a detector with smaller pixels allows for reconstructing images with higher spatial resolution at the expense of increased noise, but it also allows for reconstructing images with the same resolution and lower noise than a detector with larger pixels [41]. The PCD images show little to no overlap between the iodine and calcium signals, most clearly seen in the virtual non-contrast and virtual non-calcium images in Fig. 3 and Fig. 4. In comparison, the spectral images from the clinical scanner show significant overlap. The is explained by the well-known fact that when DECT is used to perform two-material decomposition when there are more than two materials in the object, there will be bias in the basis material estimates [4]. Although quantitative accuracy is hard to assess in a clinical specimen due to the difficulty in obtaining ground truth, the fidelity of the proposed method is validated by the agreement between the virtual non-calcium and virtual non-contrast images with known positions of vessels containing iodine. The results presented in Table 1 and Fig. 2 further validate that estimates with low bias of all three materials can be achieved. The results in Table 2 show negative noise correlation between estimates of PE and PVC as well as PVC and iodine, whereas the noise correlation between estimates of PE and iodine is positive. This is explained by the fact that the linear attenuation of iodine is best approximated by a positive part of PVC and a negative part of PE. The computed dose estimates, although approximative, indicate that it is not a large increase in dose that allows the improved differentiation of calcium and iodine. While dose reduction is one of the promises held by PCD, we refrain at this point from making any claims regarding dose savings given that the bench-top setup might not accurately reflect clinical usage. Although ring artifact correction has been applied, there are some residual ring artifacts visible in the experimental PCD images, possibly due to changes in pixel response during the time of measurement or imperfect calibration. In this particular experimental setup, the exposure times were extremely long and experiments were performed in an environment without temperature control, both circumstances increasing the sensitivity for ring artifacts. In a clinical system, both temperature and the electrical environment are highly controlled, leading to better detector stability and therefore also reduced ring artifacts. Improvements in the calibration and ring artifact correction methods could also further suppress ring artifacts.

PCDs have been used in previous studies for excised atherosclerotic plaque imaging, using both maximum-likelihood-based approaches [32] and constrained image-based approaches [27, 28]. The constrained image-based approach in [28] has also been successfully used for dual-contrast cardiac and abdominal imaging of canine models [33, 34], with both iodine and calcium as basis materials. Unconstrained three-material decomposition has also been demonstrated in previous studies [21, 26]. In this work, we demonstrate for the first time the usefulness of unconstrained three-material decomposition for imaging of anatomical samples with a silicon detector developed for high flux imaging. It is a well-known fact that single-energy CT scanners with energy-integrating detectors over-estimate plaque volume by up to 400% [42] due to calcium blooming artifacts. DECT has the potential to improve upon this, but current DECT material decomposition methods do not allow the radiologist to accurately differentiate between iodine and calcium. This makes assessing the actual diameter of the stenosis impossible, and any error in the estimation of the stenosis grade in a calcified plaque will have implications for all secondary analysis of pressure gradients and risks. The performance of current DECT material decomposition methods is expected and a well-known phenomenon within the clinical routine work where it is applied. At the same time, there is substantial evidence for the use of coronary CT-angiography as a method to provide prognostic data for low-risk patients, albeit with lower positive predictive values for a myocardial event [10]. To further enhance the predictive values, new techniques need to be developed. A clinical system based on the proposed PCD has the potential for at least two practical improvements: higher spatial resolution stemming from smaller pixels and improved separation of calcium and iodine due to more available spectral information. This could make it possible to further differentiate and characterize different plaques in order to find new risk factors. Despite recent claims to the contrary [6], our results show that it is possible to differentiate iodine and calcium with limited overlap using projection-based material decomposition with the experimental PCD. There are two important aspects of this result. The first is that the imaged object was of modest size, approximately 14 cm in diameter, which allows a high enough fraction of x-rays which are close to the k-edge of iodine in energy to escape the object unattenuated, enabling iodine to be differentiated from calcium. The second is that the spectral response of the experimental PCD preserves this information. Compared to other proposed PCD systems [6, 14‐18], the system we propose has a technical advantage that allows for handling high count rates without degrading the energy resolution and spectral performance of the detector. The PCD used in this study has been shown to have an approximately linear response for incident fluences up to 90 Mcps/mm² and a count loss of 30% at a fluence of 485 Mcps/mm² [23], with an estimated degradation of energy resolution of 0.6 keV and 3.4 keV at the respective fluences, based on measurements performed on a previous prototype [24].

A limitation of this study is that neither the focal spot size nor the tube current used in the bench-top setup is representative of clinical scans. A larger focal spot would reduce the achievable spatial resolution, and consequently, although the presented results are in line with the fact that improved spatial resolution is one of the promises of PCDs, further investigation is needed to determine how much of the presented resolution improvement can be retained in a clinical system. Also, differentiating iodine and calcium becomes more challenging with large patient anatomies [5]. Since the diameter of the imaged objects, 14 cm, is smaller than most patients, the results of this paper are not directly transferable to cardiac imaging in general without further study, but likely to other imaging areas where the anatomy is comparable in size to that of the heart phantoms, for instance neck- or pediatric imaging, where typical diameters are in ranges between 11 and 16 cm [43, 44]. Another limitation of this study was that only two heart samples were used. The data from the second heart sample showed similar results as the first but showed signs that the contrast agent had leaked between image acquisitions on the clinical scanner and in the bench-top setup, making a comparison difficult. Another limitation was that the generated basis images did not correspond exactly to water and calcium, since PE and PVC were used as basis materials. These materials were selected because they allow representing any material in the human body with positive coefficients. In the future, the calibration could be supplemented with additional measurements of water and hydroxyapatite, which would allow generating images more accurately representing water and calcium by performing a basis change from PE and PVC to water and calcium.

In conclusion, in this study, we have demonstrated that an experimental photon-counting silicon detector allows performing unconstrained three-material decomposition, thereby giving better separation of iodine and calcium than DECT. Further investigation is needed to determine how much of this improvement can be retained in a clinical system, but considering its previously shown count rate capability, these results show promise for future integration of this detector in a clinical CT scanner.