European Nuclear Medicine Guide
In 1917, the Austrian mathematician Radon formulated his theory about the possibility of reconstructing a two-dimensional function starting from its line integrals. It took more than fifty years to get to the first prototype of CT equipment, as we know by Hounsfield. After a first implementation of reconstruction with iterative algebraic algorithms, the filtered back projection (FBP) dominated the scene for 40 years, with some variations introduced to cope with spiral acquisition (interpolation of raw data) and multislice (three dimensional back projection). Since 2009, we have witnessed the introduction and rapid implementation of new iterative algorithms.
In CT equipment, for each angle of the x-ray tube and for each sampling of the detector data, attenuation profiles are acquired. The complete set of attenuation profiles obtained by a complete rotation of the detector tube assembly represented in an image is what is called the sinogram. The superposition of the projections of the attenuation profiles along the acquisition directions can give a rough first approximation of the tomographic plane with evident noise (where constant pixel values are expected they oscillate around an average value), artifacts (presence of signal different from reality), and spread of the signal. In order to partially overcome this problem, a filtering of the attenuation profiles is performed which is a convolution product with a filter function (kernel) accentuating the edges of the profile. In this way, there will be a greater definition of the object in the filtered back projection and a lower dispersion of the signal in the surrounding regions.
An important aspect of the filtered back projection is that the result we obtain depends strongly on the choice of the convolution filter. Depending on the case, we can favour noise reduction at the expense of spatial resolution (with a standard or even smoothing filter) or favour spatial resolution by accepting a higher level of noise (with an edge enhancing filter). With the filtered back-projection we obtain an approximation of the balanced reality in terms of resolution and noise.
From a mathematical point of view, the filtered reconstruction process can also be seen in the Fourier space. In the Fourier formulation of the filtered back projection, the convolution filtering becomes a ramp filtering in the frequency domain.
In the first decade of this century, the technological development of CT scans was mainly aimed at increasing the number of slices that can be acquired simultaneously with a single rotation. This is particularly useful for cardiac examinations. The number of slices progressively increased from 4 to 16, then 64, and further to 256 layers and over a volume of length 16 cm and with rotation times of the detector tube assembly of less than 0.3 s. Some tomographs have 2 x-ray tubes and two arcs of detectors placed at 90° to improve temporal resolution. Once these performances were achieved, the research focused more on the possibilities of improving image quality and reducing the dose.
Figure 1. (a) Schematic view of a third-generation tube detector assembly, common to all present CT equipment, with four rows of detectors. The number of detectors along the arc is of the order of one thousand. The nominal slice and radiation beam thickness are defined at the rotational axis distance, that is usually about 0.5 m from the focal spot. As a consequence, a slice thickness of 0.6 mm corresponds to a detector thickness along z axis of about 1.2 mm. (b) Representation of a spiral acquisition with the trajectory of the four banks of detectors around the patient.
Work has been done on increasing the performances of the detectors in terms of efficiency and response times. One manufacturer has developed a double-layer detector that discriminates between the energy of incident photons to create double-energy images. Photon count detectors are also being studied, even if important technological limits still need to be overcome before these can be included in clinical practice. From the point of view of the X-ray tubes, the range of kV available has expanded, which now ranges from 60 to 140 kV with indications on the preferred values to be used for the particular clinical application. Higher filtrations and small focal spots are available, even for high anodic currents.
The two elements that have had the greatest impact on dose reduction in CT diagnostics are the anodic current modulation systems and iterative algorithms, and these will be discussed in the following paragraphs[1].
With regard to the TC equipment used in the hybrid SPECT/CT and PET/CT machines, the number of slices commonly used is between 2 and 64 on a maximum scan length per rotation of 4 cm. A manufacturer offered a CT machine with cone beam for a period of time, but it is currently no longer in production. The kV values are between 80 and 140, and the maximum anode current can exceed 800 mA. In some cases, explicitly “non-diagnostic” CTs are implemented, in the sense that their function is limited to use for attenuation correction, scatter modelling, volume definition, and for the anatomical location of the radiopharmaceutical, although it cannot be used autonomously as CT. In this case the anodic current values are limited to some tens of mA.
Two different dose indicators, specifically defined for computed tomography, are normally used in order to compare different protocols and to assess diagnostic reference levels. The computed tomography dose indicator (CTDI) is a local dose indicator which quantifies the absorbed dose in a standard phantom (with a diameter of 32 cm for body scans and of 16 cm for head scans) for contiguous axial scans or helical scan. It intrinsically accounts for primary beam and scatter contributions. The dose length product (DLP) accounts for both the local absorbed dose and the extension of the acquired volume. It can be obtained in most cases simply by multiplying the CTDI by the scan length[2]. Both CTDI and DLP are displayed together with the other exposure parameters of each selectable acquisition protocols. Typical values of CTDI for diagnostic CT are about 60 mGy in the head region and about 10-15 mGy in the body district. Correspondent DLP values are of the order of 1000 mGy cm for a skull acquisition and of 400-600 mGy cm for a chest or abdominal scan. Regarding the CT acquisitions functional to the nuclear medicine methods, several studies have been published with a wide range of values for the CTDI and DLP indicators, which in any case are typically less than half of the values reported for diagnostic CT[3–6].
It must be clear that the CTDI indicator, despite having an undisputed utility in the definition and comparison of the acquisition protocols and being referred to cylindrical phantoms of standard size, does not correctly represent the absorbed dose of the patient’s irradiated organs which depend greatly on their actual anatomical dimensions[7]. In order to compare the relative radiation risks of the CT acquisition and of the radiopharmaceutical, effective dose is typically employed. Effective dose for CT examinations can be roughly evaluated by means of conversion factors applied to DLP values[8], or more properly with greater accuracy by dedicated software[9–11].
In the latest generation of TC equipment, an automatic modulation system of the x-ray tube current is available which, on the basis of a pre-set quality index or a nominal value of mA, modifies the intensity of the beam during the acquisition so as to maintain the image quality constant with a reduction of the dose to the patient [12,13].
There are several current modulation techniques:
modulation along the z axis: the current in the tube is modulated rotation by rotation taking into account the variation of the beam attenuation along the axis z of the patient (Fig. 2a);
angular or xy modulation: the modulation takes place during each single rotation, depending on the different attenuation of the beam intensity resulting from the elliptical anatomical conformation of the body section crossed. At shoulder level, for example, the beam is much less attenuated in the anterior-posterior direction than the lateral-lateral one. As a consequence, it is possible to reduce the current in the AP projection, thus decreasing the dose to the patient without significantly compromising the image quality (Fig. 2b);
ZXY modulation: it is given by a combination of the two modulation techniques described that intervene simultaneously (Fig. 2c).
Organ modulation: allows a reduction in the intensity of the beam corresponding to selected organs placed anteriorly, such as the crystalline lens, the thyroid, or the sinus.
Timed modulation for cardiac techniques: in cardiac CT there are different modulation techniques that intervene on the basis of cardiac cycle gating in order to reduce or completely reduce the intensity of the beam at the least useful projections for the reconstruction acquired when the organ is moving.
The presence of automatic modulation alters all known relationships between the dose and other exposure parameters in presence of a constant anodic current, so it is important to consider the behaviour of the system when other scanning parameters such as kV, the pitch, or the combinations of multi-bank detectors are modified with respect to the basic protocol. This information is essential for the adoption of optimization strategies that take into account all possible variables[14].
Automatic current modulation is particularly useful in hybrid scanners employed for both paediatric and adult patients, where a proper definition of the needed level of noise for each age class results in a significant dose reduction and image quality optimization [3].
Figure 2. Different types of current modulation obtained with an anthropomorphic phantom. (a) Z axis modulation: it is evident the peak in correspondence at the shoulders, a decrease in the chest region, and relatively higher values in the upper abdomen. (b) angular modulation: the maximum values are constant, and the minimum value depends on the relative difference between the anterior and the lateral thicknesses. (c) XYZ tube current modulation.
The benefits associated with the new generation iterative methods are the following [15]:
image noise reduction maintains the spatial resolution (in some cases improving it), unlike what happens when choosing different kernels in the FBP, where a noise reduction is paid in terms of loss of spatial resolution;
limitation of the impact of some image artifacts typically associated with the FBP reconstruction process, such as known beam hardening problems;
possible achievement of a dose reduction with different strategies (for example with smaller intensity beams with minor anodic currents or reducing the number of projections);
use of fine focus and lower kV values, even with large-sized patients, because the process no longer requires high values for the associated tube current.
About the disadvantages, it is important to highlight:
a possible increase in calculation time, but this is currently significant only for the most advanced model-based algorithms for which it could take up to tens of minutes for the reconstruction of a series of images. On the other hand, in most commercial algorithms the time increment is a few tens of seconds per series.
there is an alteration of the edges and the appearance of the image, in particular for low and medium contrasts. Images appear sometimes “plasticky” and “pixelated”.
in some cases, there may be slight, but significant, morphological changes. Particular attention should be paid to low contrast detail situations, such as abdominal CT. Several studies show that the dose reduction achievable with iterative algorithms for abdominal CT, with constant noise and contrast noise ratio, does not result in a maintenance of the detection sensitivity of low contrast details[16,17].
In commercial algorithms, the structure of the implementation is not generally known. In most cases, it is a black box where only the final effects can be evaluated. A possible classification is to distinguish the algorithms in which the iterative cycle occurs after the filtered back projection exclusively in the image domain from those in which the iterative cycle is implemented in the domain of raw data (sinograms) and then eventually also in the image, and finally from those in which the iterative cycle concerns the entire reconstruction process with multiple comparisons between forward projection and back projection[18].
The iterative algorithms in the image domain exploit denoising algorithms characterized by noise reduction capabilities while preserving spatial resolution. An important point is that they should use additional information related to the acquisition process, such as photon statistics on the directions of the different views. Otherwise, they should not be classified as iterative reconstruction algorithms but rather as post processing filters.
Methods working in the sinogram domain apply a noise reduction filter with resolution maintenance directly on the raw data, considering Poissonian statistics and then applying a filter with higher intensity (such as with a wider smoothing kernel) in the data of lower intensity (greater attenuation) and of lower intensity on the data related to projections with less attenuation. It can subsequently be reconstructed with FBP or iterative. The effect on spatial resolution and on detectability is non-linear and not entirely predictable[19].
The algorithms that operate in the entire reconstruction process are based on modelling the calculation of the attenuation profiles starting from the current image (the so-called forward projection), and through the comparison with the real attenuation profiles they calculate the corrections to be applied iteratively to the image (Fig. 3). The question in this case is what we consider in the calculation of the attenuation profiles and in the comparison with the reals. We talk about a statistical method, when we consider the contributions of Poisson and electronic noise in the definition of the calculated profiles, basically with algebraic methods, and in calculating the corrections we will attribute a lower weight to the projections affected by high noise[20]. Instead, we are talking about model-based methods, when we consider the different real aspects of the physical process of signal generation and acquisition, e.g. finite size of the focal spot, of the voxel and of the detectors, the contributions of the scatter, and so on. In this second case, a world of different possibilities opens. Obviously, more elements of the real world are actually considered, and more the calculation process becomes complex and long. To improve the spatial resolution, different methods of discretization can be considered subtler than the volume of voxels to be calculated, for example by halving the linear dimension of the voxels, or considering density blob functions. For beam components, it is possible to switch from single lines to multiple lines or parallel bands or diverging bands and taking into account the focal spot size and detectors. Increasing the complexity, algorithms have been developed to model scatter radiation using analytical methods or Monte Carlo so the effects related to the energy spectrum can be considered, for example correcting the artefacts from hardening of the beam [21]. The performances of the model based algorithms are significantly better than those obtained with statistical algorithms, as shown by some comparative studies [22,23].
The iterative algorithms in CT find important applications in the particular situations indicated by the terms data sparsity and compressed sensing, in other words, the situations where there is a lack of acquisition data with a regular structure (e.g. reduction of the number of views) or with unstructured mode (e.g. partial rotation angle, absence of data to the detector on some views)[24].
Figure 3. Schematic flow chart of an iterative algorithm operating in the entire reconstruction process, with multiple forward and backward projections. When the comparison between the calculated and the acquired profiles gives differences below a predefined threshold, or after a maximum number of iterations, the final image is achieved.
Common to most commercial iterative algorithms is the possibility to choose the intensity level of the iterative, so that the final image can be defined as a linear combination of the FBP image and the image with the maximum contribution of the iterative. For some producers, there is talk of an iterative percentage, for others, of strength or level. During the installation phase, the application specialist typically proposes initial iterative levels for the different protocols, and afterwards, it is possible to optimize the iterative level [25]. In nuclear medicine equipment, where the main purpose of CT images are attenuation corrections and anatomical localization, high levels of iterative reconstruction contributions with significant dose reductions are applicable [4–10].
The accuracy of CT images is often affected by the presence of artefacts of different origins, such as noise, beam hardening, scatter, motion, cone beam, ring, and metal artefacts. These last are the most common, and they are caused by metallic implants within a patient’s body: dental fillings, orthopaedic prosthesis, pacemakers, and cardiac defibrillators. The high attenuation of the x-ray beam across these devices determines a gap in the projection data, a loss of the information of the actual attenuation of the structures close to the metallic implants, and the resulting appearance of streaking artifacts and void regions on the reconstructed CT image [26].
The use of CT images with streaking artefacts in CT-based attenuation correction of PET data can cause a propagation of the artefacts into the corresponding PET images in the form of overestimated or under-estimated activity concentration regions which might lead to a false diagnosis of the disease.
Also, in diagnostic radiology and in radiotherapy, the presence of metal artefacts can lead to substantial errors. For these reasons, several metal artefact reduction (MAR) strategies have been developed in order to limit the impact of these artifacts and to improve the accuracy of density values in the regions surrounding metallic elements. They can be divided into implicit methods, with the choice of the best exposure parameters to limit the impact of the metallic element, and explicit approaches, using software reconstruction and processing algorithms. Implicit approaches include gantry tilt to exclude the metallic object from the tomographic plane as well as the increase of beam energy and intensity. The most common methods belong to the explicit category, with various techniques either in the sinogram or image domains. A recent review on a broad range of MAR methods showed that the sinogram-based methods seem, in general, more accurate than imaged-based.
The introduction of deep learning algorithms in computed tomography image reconstruction can serve several purposes. Similar to previous iterative algorithms, deep learning systems offer the possibility of reducing image noise and consequently allowing a reduction in patient dose while maintaining the information of interest. This type of application can be implemented in practice by having the neural network intervene at different levels [27]:
1. One can speak of 'direct' DL reconstruction when, starting from the low-dose sinogram, the neural network directly reconstructs the final tomographic image, no longer using the classical FBP;
2. Algorithms with neural networks that act solely on the synogram to optimise the collected raw data are referred to as 'indirect synogram-based';
3. Algorithms with neural networks acting only in image space can be referred to as 'indirect image based';
4. Algorithms that combine the two applications of neural networks mentioned above, while retaining the basis of FBP or IR reconstruction, are referred to as 'indirect hybrid' reconstruction.
In the case of 'indirect image based', there are commercial AI system products independent of the equipment manufacturer that can be applied directly to any type of tomographic image. In all other cases, the solutions are implemented directly by the CT manufacturers and the type should be stated in the technical specifications of the product.
Similarly to iterative algorithms, the actual performance of these algorithms needs to be properly validated. In terms of dose and noise reduction, they have been shown to be significantly superior to iterative algorithms. However, the nature of neural network operation poses significant risks in terms of possible alterations in image details with an impact on diagnosis, so it is crucial to investigate these possible behaviours already in the training and testing phases of neural networks. Medical physicists are developing specific metrics to quantify the retention of information in CT images reconstructed with AI, particularly in cases of special requirements for low-contrast detectability or high spatial resolution.
An interesting application of AI in improving image quality concerns the possible removal of the scatter contribution from tomographies, which is particularly useful in the case of CBCT where scatter removal grid systems are less efficient than those found in MSCT tomographies. Methods based on Monte Carlo simulations take several hours to estimate and remove the scatter contribution from a tomographic data set. With a deep learning network, estimation of the scatter contribution can be done in 3 ms for a CT view and 10 ms for a CBCT projection. This provides a fast and accurate alternative to MC simulation that can be used for geometries and beam qualities other than those used in training [28].
In the field of cardiac CT, algorithms have been developed that are able to evaluate and compensate for loss of definition due to motion [29]. Even for metal artefacts there are deep learning algorithms with similar or in some cases higher performance than iterative algorithms. [30].
A photon counting CT is a CT with a detector capable of measuring the energy deposited by a single incident photon and consequently 'counting' the individual photons and grouping them according to their relative energy. Photon counting CTs use the electronic technology of direct conversion, whereby the incident photon in a semiconductor material directly produces an electrical signal that can be measured by the creation of an electron-gap pair. The charges generated in this way move perpendicular to the incident surface of the detector and thus enable high spatial resolution to be achieved. In conventional scintillation detectors, also called Energy Integrated Detectors (EID) after the introduction of photon counting CT, it is necessary to use separation baffles between the cells of the semiconductor material in order to achieve adequate resolution. For the single event, the signal collection time is approximately 100 times longer for EID detectors than for PCCTs (2500 to 25 ns). PCCT detectors therefore offer the possibility of obtaining images with significantly higher spatial resolution for the same radiant dose and noise, or of reducing the noise (and consequently the dose) for the same spatial resolution.
Particularly appreciated are the advantages of this technology for examinations of patients with cardiovascular and pulmonary diseases with high resolution, but also in oncological follow-up, orthopaedic, neurological and paediatric examinations [31]. From the point of view of the spectral separation of the measured signal, it is possible to use multiple energy channels in contrast to classical dual energy CT scans. This possibility opens up possible applications in the separation of the content of different materials and the use of several simultaneously distinguishable contrast agents by means of spectral imaging [32].
The actual ability to separate the different incident photons in energy and thus the resulting spectral imaging is, however, still the subject of research and development to overcome certain technological limitations. The following physical phenomena have a significant impact on these limits:
Pulse pile-up: the arrival of two consecutive photons in a time interval shorter than the duration of the single reading leads to the summation of the signal deposited by the two photons and the possible interpretation of a single higher-energy photon;
Charge sharing: the collection of electrical charges in the semiconductor material of the detector in a region at the boundary between two electrodes of the array of the readout electronics can lead to an underestimation of the signal amplitude of the incident photon;
K fluorescence: in the case of energy transitions between orbitals of the detector atoms, fluorescence photons can be emitted, which in turn will generate spurious signals.
The accuracy of spectral imaging and evaluations of the relative contribution of materials of interest (such as iodine or calcium) can be assessed with appropriate test phantoms having inserts with known concentrations of these materials [33].
Manufacturers are developing PCCT detectors with two different approaches, called 'face on' and 'edge on'. The face-on approach has a detector with a sensitive layer arranged perpendicular to the incident photons, with materials such as cadmium telluride (CdTe) or cadmium zinc telluride (CZT). For edge-on, on the other hand, silicon is used with layers arranged parallel to the incident photons and a total thickness of a few cm. The first PCCT tomographs to come onto the market are all face-on. Edge-on systems are promising in some aspects such as a possible better energy resolution and a lower probability of k-fluorescence events. On the other hand, they also have possible disadvantages related to a higher contribution of Compton scattering within the detector and an overall more complex realisation system.
1. McCollough CH, Chen GH, Kalender W, Leng S, Samei E, Taguchi K, et al. Achieving Routine Submillisievert CT Scanning: Report from the Summit on Management of Radiation Dose in CT. Radiology 2012;264:567–80. https://doi.org/10.1148/radiol.12112265.
2. Quality Assurance Programme for Computed Tomography: Diagnostic and Therapy Applications 2016. https://www.iaea.org/publications/8751/quality-assurance-programme-for-computed-tomography-diagnostic-and-therapy-applications (accessed July 29, 2020).
3. Yang C-C, Yang B-H, Tu C-Y, Wu T-H, Liu S-H. Low-dose computed tomography scans with automatic exposure control for patients of different ages undergoing cardiac PET/CT and SPECT/CT. Nucl Med Commun 2017;38:546–55. https://doi.org/10.1097/MNM.0000000000000679.
4. Grosser OS, Kupitz D, Ruf J, Czuczwara D, Steffen IG, Furth C, et al. Optimization of SPECT-CT Hybrid Imaging Using Iterative Image Reconstruction for Low-Dose CT: A Phantom Study. PLoS ONE 2015;10:e0138658. https://doi.org/10.1371/journal.pone.0138658.
5. Sibille L, Chambert B, Alonso S, Barrau C, D’Estanque E, Tabaa YA, et al. Impact of the Adaptive Statistical Iterative Reconstruction Technique on Radiation Dose and Image Quality in Bone SPECT/CT. J Nucl Med 2016;57:1091–5. https://doi.org/10.2967/jnumed.115.164772.
6. Gupta SK, Trethewey S, Brooker B, Rutherford N, Diffey J, Viswanathan S, et al. Radionuclide bone scan SPECT-CT: lowering the dose of CT significantly reduces radiation dose without impacting CT image quality. Am J Nucl Med Mol Imaging 2017;7:63–73.28533938
7. McCollough CH, Leng S, Yu L, Cody DD, Boone JM, McNitt-Gray MF. CT dose index and patient dose: they are not the same thing. Radiology 2011;259:311–6. https://doi.org/10.1148/radiol.11101800.
8. Deak PD, Smal Y, Kalender WA. Multisection CT protocols: sex- and age-specific conversion factors used to determine effective dose from dose-length product. Radiology 2010;257:158–66. https://doi.org/10.1148/radiol.10100047.
9. impactscan.org | ctdosimetry.xls - ImPACT’s ct dosimetry tool n.d. http://www.impactscan.org/ctdosimetry.htm (accessed July 29, 2020).
10. Stamm G, Nagel HD. [CT-expo--a novel program for dose evaluation in CT]. Rofo 2002;174:1570–6. https://doi.org/10.1055/s-2002-35937.
11. Ding A, Gao Y, Liu H, Caracappa PF, Long DJ, Bolch WE, et al. VirtualDose: a software for reporting organ doses from CT for adult and pediatric patients. Phys Med Biol 2015;60:5601–25. https://doi.org/10.1088/0031-9155/60/14/5601.
12. Keat, Nicholas. (2005). CT scanner automatic exposure control systems. 10.13140/RG.2.1.3128.4720.
13. Lee CH, Goo JM, Ye HJ, Ye S-J, Park CM, Chun EJ, et al. Radiation dose modulation techniques in the multidetector CT era: from basics to practice. Radiographics 2008;28:1451–9. https://doi.org/10.1148/rg.285075075.
14. Rampado O, Marchisio F, Izzo A, Garelli E, Bianchi CC, Gandini G, et al. Effective dose and image quality evaluations of an automatic CT tube current modulation system with an anthropomorphic phantom. European Journal of Radiology 2009;72:181–7. https://doi.org/10.1016/j.ejrad.2008.06.027.
15. Beister M, Kolditz D, Kalender WA. Iterative reconstruction methods in X-ray CT. Phys Med 2012;28:94–108. https://doi.org/10.1016/j.ejmp.2012.01.003.
16. Schindera ST, Odedra D, Raza SA, Kim TK, Jang H-J, Szucs-Farkas Z, et al. Iterative reconstruction algorithm for CT: can radiation dose be decreased while low-contrast detectability is preserved? Radiology 2013;269:511–8. https://doi.org/10.1148/radiol.13122349.
17. Goenka AH, Herts BR, Obuchowski NA, Primak AN, Dong F, Karim W, et al. Effect of Reduced Radiation Exposure and Iterative Reconstruction on Detection of Low-Contrast Low-Attenuation Lesions in an Anthropomorphic Liver Phantom: An 18-Reader Study. Radiology 2014;272:154–63. https://doi.org/10.1148/radiol.14131928.
18. Willemink MJ, de Jong PA, Leiner T, de Heer LM, Nievelstein RAJ, Budde RPJ, et al. Iterative reconstruction techniques for computed tomography Part 1: technical principles. Eur Radiol 2013;23:1623–31. https://doi.org/10.1007/s00330-012-2765-y.
19. Kachelrieß M, Watzke O, Kalender WA. Generalized multi-dimensional adaptive filtering for conventional and spiral single-slice, multi-slice, and cone-beam CT. Medical Physics 2001;28:475–90. https://doi.org/10.1118/1.1358303.
20. Xu J, Tsui BMW. Electronic Noise Modeling in Statistical Iterative Reconstruction. IEEE Trans Image Process 2009;18:1228–38. https://doi.org/10.1109/TIP.2009.2017139.
21. Nuyts J, De Man B, Fessler JA, Zbijewski W, Beekman FJ. Modelling the physics in the iterative reconstruction for transmission computed tomography. Phys Med Biol 2013;58:R63-96. https://doi.org/10.1088/0031-9155/58/12/R63.
22. Löve A, Olsson M-L, Siemund R, Stålhammar F, Björkman-Burtscher IM, Söderberg M. Six iterative reconstruction algorithms in brain CT: a phantom study on image quality at different radiation dose levels. Br J Radiol 2013;86. https://doi.org/10.1259/bjr.20130388.
23. Miéville FA, Gudinchet F, Brunelle F, Bochud FO, Verdun FR. Iterative reconstruction methods in two different MDCT scanners: physical metrics and 4-alternative forced-choice detectability experiments--a phantom approach. Phys Med 2013;29:99–110. https://doi.org/10.1016/j.ejmp.2011.12.004.
24. Yu H, Wang G. Compressed sensing based interior tomography. Phys Med Biol 2009;54:2791–805. https://doi.org/10.1088/0031-9155/54/9/014.
25. Rampado O, Busso S, Garabello D, Marengo E, Valerio M, Capello S, et al. Aortic CT angiography dose reduction: investigation of optimal noise index and iterative algorithm strength in combination with low kV. Radiol Med 2016;121:291–300. https://doi.org/10.1007/s11547-015-0611-4.
26. Abdoli M, Dierckx RAJO, Zaidi H. Metal artifact reduction strategies for improved attenuation correction in hybrid PET/CT imaging. Medical Physics 2012;39:3343–60. https://doi.org/10.1118/1.4709599.
27. Lennart R Koetzier, Domenico Mastrodicasa, Timothy P Szczykutowicz, Niels R van der Werf, Adam S Wang, Veit Sandfort, Aart J van der Molen, Dominik Fleischmann, Martin J Willemink Deep Learning Image Reconstruction for CT: Technical Principles and Clinical Prospects Radiology. 2023 Mar;306(3):e221257. doi: 10.1148/radiol.221257
28. Joscha Maier, Elias Eulig, Tim Vöth, Michael Knaup, Jan Kuntz, Stefan Sawall, Marc Kachelrieß Real-time scatter estimation for medical CT using the deep scatter estimation: Method and robustness analysis with respect to different anatomies, dose levels, tube voltages, and data truncation Med Phys. 2019 Jan;46(1):238-249. doi: 10.1002/mp.13274
29. Joscha Maier, Sergej Lebedev, Julien Erath, Elias Eulig, Stefan Sawall, Eric Fournié, Karl Stierstorfer, Michael Lell, Marc Kachelrieß Deep learning-based coronary artery motion estimation and compensation for short-scan cardiac CT Med Phys. 2021 Jul;48(7):3559-3571. doi: 10.1002/mp.14927.
30. Mark Selles, Jochen A C van Osch, Mario Maas, Martijn F Boomsma, Ruud H H Wellenberg Advances in metal artifact reduction in CT images: A review of traditional and novel metal artifact reduction techniques Eur J Radiol. 2024 Jan:170:111276. doi: 10.1016/j.ejrad.2023.111276.
31. Philippe C Douek, Sara Boccalini, Edwin H G Oei, David P Cormode, Amir Pourmorteza, Loic Boussel, Salim A Si-Mohamed, Ricardo P J Budde Clinical Applications of Photon-counting CT: A Review of Pioneer Studies and a Glimpse into the Future Radiology. 2023 Oct;309(1):e222432. doi: 10.1148/radiol.222432
32. Willemink MJ, Persson M, Pourmorteza A, Pelc NJ, Fleischmann D. Photon-counting CT: Technical Principles and Clinical Prospects. Radiology. 2018 Nov;289(2):293-312. doi: 10.1148/radiol.2018172656.
33. Stevan Vrbaski, Steve Bache, Jayasai Rajagopal, Ehsan Samei Quantitative performance of photon-counting CT at low dose: Virtual monochromatic imaging and iodine quantification doi.org/10.1002/mp.16583