Iterative reconstruction in CT imaging; Image quality and radiation doses

(1)

Iterative reconstruction in CT imaging;

Image quality and radiation doses

Dissertation for the degree of PhD April 2020

Kristin Jensen

Department of Diagnostic Physics Division of Radiology and Nuclear Medicine

The Intervention Centre

Division of Emergencies and Critical Care

Department of Physics

Faculty of Mathematics and Natural Sciences

(2)

(3)

i

Acknowledgement

This work has been carried out at the Department of Diagnostic Physics and the Intervention Centre at Oslo University Hospital and the Department of Physics at University of Oslo.

I would like to thank my main supervisor, Anne Catrine Martinsen, for facilitating the opportunity to start and complete this PhD project. Without you this would not have been possible. Thank you for all the ideas, valuable feedback, discussions and support. Even though your schedule is more than full, you have always had time when I have needed help. Your optimism is invigorating and the engagement and initiative you have had for medical physicists and physics in radiology are impressive. It has been interesting to follow the tasks you have decided to pursue, and now you will pursue something quite different. Good luck and the best wishes!

Thank you to my co-supervisors, Anders Tingberg and Erik Fosse. Your different insights and experiences in medical technology, imaging and research are valuable, and I have appreciated your feedback both on content, structure and writing. You have also been encouraging and supportive all the way, and I have always felt uplifted after meetings. Also, a big thank you to Örjan Smedby who introduced me to visual grading regression and STATA, and to Atle Bjørnerud who stepped in as supervisor at the last minute.

In addition to my supervisors I have had help with scanning phantoms and writing articles. To have access to firsthand knowledge on CT scanners, anatomy and the radiology field makes my day more interesting. Thank you to all my co-authors and scanning help. Ingrid Helen Hauge and Øystein Bech Gadmar, thank you for proofreading of articles and this dissertation. Your comments have been very helpful.

Lunch and coffee breaks are important for thinking, and I am very grateful for having so many nice colleagues always making these breaks fun and enjoyable. A special thanks to Hilde Kjernlie Andersen for facilitating my workday so I could finish this thesis, and to all the physicists at the Department of diagnostic physics. No matter what problems there are, you are always willing to help, and I am grateful to work with you.

Thank you to my parents and brother who made my childhood safe, yet still eventful, and gave me the prerequisites to utilize the opportunities brought by life. And to my closest, Marius, Joakim and Hedda; I look forward to every day with you.

(4)

ii

(5)

iii

Thesis summary

The purpose of this thesis is to evaluate different iterative reconstruction techniques with respect to image quality and radiation dose dependency across CT scanners from all major vendors. The evaluation includes how traditional quantitative measurements correlate with visual grading analysis and receiver operating characteristics when images are reconstructed with iterative reconstruction.

Two anthropomorphic image quality phantoms - liver and chest - were imaged with different CT scanners at different dose levels. Images were reconstructed with filtered back projection and different iterative reconstruction algorithms. Images of the liver phantom were analysed using receiver operating characteristics, visual grading regression and the following quantitative image properties: Standard deviation of Hounsfield units (image noise), signal to noise ratio and contrast to noise ratio. Images of the chest phantom were analysed using visual grading regression and image texture analysis by noise power spectrum.

All quantitative measurements were improved using iterative reconstruction both in abdomen and chest. Receiver operating characteristics showed that some algorithms improved liver lesion detection at low doses, while other worked better at higher doses. Some did not change

detection while one decreased detection. Visual grading analysis showed that the decreased detection may be because of artefacts or degraded sharpness. Evaluated anatomical structures and image quality properties in the chest were improved or the same as with filtered back projection with a few exceptions. One algorithm maintained the noise power spectrum while the other algorithms shifted the spectrum towards lower frequencies in varying degree.

Performance of eight iterative reconstruction algorithms differs across vendors and with image quality parameters or evaluation criteria, dose level and reconstruction kernel. Traditional measurements like image noise, signal to noise ratio and contrast to noise ratio are not as useful in images reconstructed with iterative reconstruction as in images reconstructed with filtered back projection.

List of papers

1. Comparing five different iterative reconstruction algorithms for computed tomography in an ROC study. Kristin Jensen, Anne Catrine T. Martinsen, Anders Tingberg, Trond Mogens Aaløkken and Erik Fosse. European Radiology (2014)

2. Image quality in oncologic chest computerized tomography with iterative reconstruction.

A phantom study. Kristin Jensen, Trond Mogens Aaløkken, Anders Tingberg, Erik Fosse, Anne Catrine T. Martinsen. Journal of computer assisted tomography (2016)

3. Quantitative measurements versus receiver operating characteristics and visual grading regression in CT images reconstructed with iterative reconstruction. Kristin Jensen, Hilde Kjernlie Andersen, Örjan Smedby, Bjørn Helge Østerås, Anette Aarsnes, Anders Tingberg, Erik Fosse, Anne Catrine Martinsen. Academic Radiology (2017)

4. Evaluation of image quality for seven iterative reconstruction algorithms in chest CT imaging; a phantom study. Kristin Jensen, Guro Hagemo, Anders Tingberg, Claudius Steinfeldt-Reisse, Georg Karl Mynarek, Rodriguez Jezabel Rivero, Erik Fosse, Anne Catrine Martinsen. Journal of computer assisted tomography (2020)

(6)

iv

(7)

v

1 Background

Computed Tomography CT was introduced in the early 1970s by Godfrey Hounsfield and Allan Cormack, making it possible to obtain axial patient images (1-3). The technology underwent a rapid development from one detector to several detectors covering the width of the patient, from left to right, with multiple slices. Today, CT is widely used for diagnosis in most radiological areas;

for example, musculoskeletal, oncology, trauma and stroke. The main advantages are three- dimensional images and fast volume coverage, making it possible to get high resolution images from neck to pelvis in a couple of seconds. One disadvantage of CT imaging is the radiation dose delivered to the patient. Vendors have developed techniques to keep the radiation dose to a minimum while preserving image quality; these include improvements in detector and x-ray tube technology, dose modulation, and image reconstruction. The image reconstruction method used during the first 30 years of using CT, was filtered back projection. This method required only moderate computational power and was the main method until iterative reconstruction

algorithms became commercially available in 2008 (4). Iterative reconstruction algorithms reduce image noise, but they also shift the noise content to lower spatial frequencies, changing the texture of the images. Artificial intelligence (AI) based reconstruction algorithms were introduced in 2018. These algorithms will have the appearance of the image data set they are trained on (5).

1.1 Computed Tomography

In CT, x-rays from an x-ray tube are transmitted through the patient. Detectors on the opposite side of the gantry detect the remaining x-ray photons and these signals give information on the attenuation of different tissues in the body (Figure 1). To receive enough information about details inside the body, the tube and detectors rotate around the patient, collecting information 1000 5000 times in one rotation (6). All these data, which are called raw data or sinogram data, are reconstructed to an image which can then be interpreted to answer the objective of the patient examination. This process is called the reconstruction process.

Figure 1. Computed tomography, CT. Image of a modern CT scanner (left) and schematic (right).

(10)

2

1.2 Image reconstruction in CT

When x-ray photons are detected by the rotating detectors, the signal appears like a collection of sinus curves, hence the term sinogram data. In order to obtain an image that resembles the patient, the sinogram data must be processed by mathematical algorithms. The first CT scanners used iterative reconstruction (algebraic reconstruction technique ART) to process the data (7, 8). As the number of detectors increased, rotation times were reduced and spatial resolution increased. Computational power became a major limitation. Therefore, an analytical

reconstruction method filtered back projection was introduced to process the data faster albeit based on simplifications. This was the main method to reconstruct CT images until iterative reconstruction algorithms were relaunched in 2008 (9).

Filtered back projection

Figure 2. Left: an object, f, in the x-y plane. Parallel lines from a source, through the object, to a detector element t at two different angles, θ. The sinogram data at the two different angles are denoted Pθ1(t) and P

θ2(t). Used with permission from Avinash C Kak and Malcolm Slaney (10). Upper right:The sinogram data of a phantom. Lower right: The reconstructed image of the sinogram data above. Both phantom images used with permission from Jiang Hsieh (11).

The sinogram signal data from the detectors, P (t), is preserved over the radon transform. This transform is a function defined on the space of straight lines in a 2D plane by the line integral along each line. To recover the scanned object from the sinogram data, the radon inversion formula or back projection formula can be used, Eq 1. From the sinogram data P (t), one wants to reconstruct the object f(x,y)(Figure 2) (12, 13):

f(x,y) = ∫ 𝐹⁻ 𝐹 𝑃 (t)} * h} d𝜃 (1)

(11)

3

For each , compute the 1D Fourier Transform, F{P (t)}, of the sinogram data P (t) Multiply F{P (t)} with a filter function, h

Compute the inverse Fourier Transform to obtain the filtered projections Integrate all results over

There are more projections passing through the center of the object compared to the periphery.

This will result in a low pass blur which can be reduced with filtering. The filter function, h, is called the clinical kernel. In abdominal imaging, soft or standard kernels are used, while in chest imaging more edge enhancing filters are used. The disadvantage of these kernels is that one must choose between smoothing kernels which smooth noise and enhance low contrast resolution, or edge enhancing kernels which sharpen edges but increase image noise (14).

Filtered back projection has the advantage of fast reconstruction times of 5.4 80 images per second (15), and if the radiation dose is not to low, the image quality is good. Still, there are limitations to the filtered back projection reconstruction technique: It does not account for quantum noise and electronic noise in the signal. It also assumes a point source and point detector, making the radiation beam a pencil beam and the imaged voxel becomes a

dimensionless point (16). Considerable image noise and artefacts will appear at low radiation doses.

Iterative reconstruction

The iterative reconstruction algorithm formulas are usually divided into a data term, H, which is a fitting model of the measured projection data, and a regularisation term, n, which incorporates the nonuniformities of the system, such as noise. This can be written

P (t) =H*µ + n, (2)

where P (t) is the sinogram data, H is the projection process, µ the attenuation coefficients and n additional noise (17).

The algorithm usually starts with an initial image, most often the filtered back projection or an image based on it. The projections of this initial image are compared to the measured projections, and corrections depending on the models used are performed on the initial data set. Then a new comparison with the measured data is performed and new corrections are made. This procedure is repeated a predetermined number of iterations, or until the desired image quality is reached (7, 14, 16).

To do this, the process is minimizing the differences between the measured data, P (t), and the initial data, or synthesised data, by the following equation (18, 19):

𝑥 arg 𝑚𝑖𝑛 𝑃 𝑡 Hμ 𝛽𝑛 , (3)

where 𝑥 is the next estimation of the reconstructed image, w is a weighting function and is a parameter controlling the strength of the regularisation.

(12)

4

The iterative reconstruction algorithms are divided into two main categories: Algebraic and statistical algorithms. Algebraic algorithms solve several linear equations (H in equation (2) and (3)) with the projection values being the sum of all attenuation values along lines through the patient. Statistical algorithms introduce a weighting (w in equation (3)) of low or high uncertainty data in the data fitting, and fitting can be achieved with for example maximum likelihood, least squares, or maximum a posteriori estimators (14, 17).

The parameter is s all set to er small or at the end of the reconstruction process, especially when used with edge-preserving algorithms, not to introduce artefacts (20).

The complexity of commercial iterative reconstruction algorithms varies. They can be statistical only, mainly reducing image noise, or they can combine this with the regularisation or edge preserving algorithms. If only the regularisation term is used, noise reduction is achieved. Other models must be incorporated into the data term, H, and these can be more or less complex.

Following is a list of a few of the possible models (17, 20):

Noise models: Noise can be random photon noise (quantum noise), anatomical noise, structural noise due to photon starvation, electronic noise and characteristic noise from the particular detector element. Since radiation is produced by a statistical process, the number of photons from the tube to the detector will vary even when exposure

parameters are identical. Poisson statistics are typically used to reflect this. Quantum noise is dominant at higher radiation doses, while electronic and structural noise are a more dominant part of the noise at lower radiation doses (21).

Radiation and detection physics: Physical modeling of photon interactions both in patient and detector. This includes photoelectric effect as well as Compton and Rayleigh

scattering. Usually a linear model is used, assuming that all photons have the same energy. The likelihood of different interaction processes is calculated. Non-linear effects like scatter and beam hardening are not accounted for. Information about the x-ray spectra and the imaged object is needed to account for these.

Object models: Utilize known information about the scanned object or desired image behaviour.

Scanner models: Model the scanner geometry and properties, for example focal spot, bowtie filtration, collimation (cone beam), the distances between source, isocenter and detector, and detector geometry.

Optics model: Models photon paths from the focal spot through the patient to the detector element. Scattering and other random variations are accounted for.

Prior object information modeling/regularisation: Smoothens pixel values which are unrealistically high or low compared to neighboring pixel values and removes negative attenuation values. This part helps the algorithms converge to images with lower noise, thus avoiding higher noise levels. This is the n term in equation (2).

The models can be implemented in the sinogram space, the image data domain or in a loop between sinogram and image data domains; see Figure 3.

(13)

5

Figure 3. Iterative reconstruction loops can work in the sinogram data, image data or between the sinogram and image data.

All algorithms have a statistical noise model to reduce image noise. Image noise is reduced but the image texture may change compared to filtered back projection. To maintain a familiar image texture, many of the algorithms mix data from the iterative process with data from filtered back projection. Adaptive algorithms reduce image noise more when it is higher. Table 1 lists the commercially available iterative reconstruction algorithms, and a short description of the algorithms follows.

Hybrid statistical algorithms

The hybrid statistical algorithms correct the sinogram data to reduce artefacts. In the image domain, a statistical noise model is used to iteratively reduce noise while preserving edges. Data from the filtered back projection is also included in the reconstruction (7, 16).

Adaptive Statistical Iterative Reconstruction, ASiR

Vendor advertised behavior: ASiR starts with the filtered back projection image. The data is compared with the originally measured data in sinogram space. Based on models of photon statistics and the scanned object, noise is reduced while edges are preserved (22). This is repeated several times, and the result is blended with the FBP image. The amount of data from filtered back projection and ASiR is dependent on the chosen level of ASiR from 0 to 100% (23).

Peer reviewed literature characterization: Quantitative measurements are improved compared to filtered back projection and lower strengths of ASiR (24-26), but pixelation is reported, especially at high strengths (27, 28). Higher strengths of ASiR will also increase the low frequency noise proportion (29).

(14)

6 Table 1. Iterative reconstruction algorithms

Type Algorithm – Launched

Vendor Mode Levels Models:

Sinogram data

Models:

Image data

Recon- struction time

Statistical, hybrid

ASiR – 2008 (22)

GE Healthcare

- 0 100% Noise model Comp-

arable to FBP iDose – 2010

(30)

Philips Healthcare

- 1 7 Noise model Anatomical

based, multi frequency noise reduction

Comp- arable to FBP

SAFIRE – 2010 (31)

Siemens Health- ineers

- 1 5 Noise model Statistical

noise regularisation

Comp- arable to FBP AIDR 3D – 2010

(Enhanced added from software version 7.0) (32)

Canon Medical Systems

- Mild,

standard, strong, enhanced

Noise model Anatomical based denoising

Comp- arable to FBP

Partial model-based ASiR-V – 2013 (19)

GE Healthcare

- 0 100% Noise model

Physics model Object model

Comp- arable to FBP IMR – 2012 (33) Philips

Healthcare

Body Routine Body soft Body SharpPlus

1 3 Noise model Scanner model Object model

Upper abdomen

5 min ADMIRE – 2013

(34)

Siemens Health- ineers

- 1 5 Noise model Noise model

System model Comp- arable to FBP

Model-based

Veo – 2010 (35) GE Healthcare

Standard 1 Noise model

Physics model Object model Scanner model Optics model

Upper abdomen

20 40 min

FIRST – 2015 (18)

Canon Medical Systems

Lung Body (Sharp) Cardiac (Sharp) Bone

Brain (LCD/CTA) Mild, standard, strong

Noise model Scanner model Optics model Cone beam model

Anatomical based noise regularisation

Upper abdomen

3 15 min

iDose

Vendor advertised behavior: iDose uses an iterative process in sinogram space to reduce noise.

Each projection is examined to find data resulting from noisy measurements. Noisy data and edges are differentiated in order to not affect the edges. Data is then propagated into image space where noise is localized. Here, a noise model is applied in order to reduce noise while preserving edges. Then a noiseless anatomical structure model corresponding to the local topology is used to further reduce noise, and a dynamic model-based noise removal is used to preserve the noise power spectrum (36, 37). Levels of iDose range from 1 to 7, and a higher level means that there is an increased reduction in image noise (23).

Peer reviewed literature characterization: Quantitative measurements are improved compared to filtered back projection (38, 39), but high levels of iDose are reported to give blotchy and

pixelated images (40).

(15)

7 Sinogram Affirmed Iterative Reconstruction, SAFIRE

Vendor advertised behavior: A weighted filtered back projection is reconstructed followed by a re-projection into sinogram space. The original data are compared to the reconstructed sinogram data, and imperfections are corrected and artefacts removed. The new sinogram data is

reconstructed with the weighted filtered back projection. This process is repeated, and in the process, a dynamic sinogram data-based noise model is applied in order to reduce noise while preserving edges. The number of iterations is dependent on the type of exam that is conducted.

Following this process, an iterative loop in image space is performed reducing noise through a statistical optimisation process applying the knowledge of how noise propagates into image space (31). Five levels of iterative strength are selectable, with quantitative measurements improving with higher levels (23).

Peer reviewed literature characterization: The noise power spectrum is shifted towards lower frequencies (41), and pixelated appearance is reported at high level settings (42, 43). A CT liver examination with filtered back projection and SAFIRE is shown in Figure 4.

Figure 4. CT liver images with filtered back projection (left column), the hybrid statistical algorithm SAFIRE 3 (upper right), and the partial model-based algorithm ADMIRE 3 (lower right).

(16)

8 Adaptive Iterative Dose Reduction 3D, AIDR 3D

Vendor advertised behavior: AIDR 3D uses a scanner model and a statistical noise model to minimize photon starvation, and electronic and statistical noise models in sinogram data to reduce random and structured noise. The algorithm is adaptive, reducing more noise at low doses compared to high doses. In image space, a voxel based anatomical model reduces quantum noise iteratively. The filtered back projection image is blended with the processed image in each iteration to maintain image texture. The number of iterations and the amount of blending with filtered back projection is dependent on anatomical region and type of examination (32). Three levels of iterative strength can be chosen; mild, standard and strong. AIDR Enhanced was added in software version 7 (44).

Peer reviewed literature characterization: As with the other hybrid, statistical algorithms, quantitative measurements are improved with AIDR compared to filtered back projection (45, 46), and the noise power spectrum is shifted to lower frequencies (37). A chest image with this algorithm is shown in Figure 5.

Partial model-based algorithms

These algorithms use more complex models than the hybrid statistical algorithms. More

information on the scanner geometry is usually implemented in the algorithms, without modelling the full optics which is very time consuming. Some of these algorithms also have more advanced models on the radiation physics.

ASiR-V

Vendor advertised behavior: ASiR-V is the next generation of ASiR, using more advanced noise and object models compared to ASiR, and adding physics modelling. ASIR-V de-emphasizes system optics modelling compared to Veo, resulting in faster reconstruction times. The noise model includes electronic and photon noise, and noise characteristics of the reconstructed images. This includes photon statistics throughout the imaging chain, and characterisation of the imaged object obtained from clinical and phantom data (19, 47). Its vendor reports dose

reductions up to 82% (19).

Peer reviewed literature characterization: Noise frequencies are shifted to lower frequencies at higher filter strengths, and the noise content is dependent on slice thickness (48, 49).

Iterative Model Reconstruction, IMR

Vendor advertised behavior: Knowledge of system geometry, x-ray statistics, object properties and desired characteristics are implemented in the IMR process to improve the image (17, 23, 33). Little information is published on the theory of this algorithm. Three strengths are available.

Peer reviewed literature characterization: Quantitative measurements are improved (23, 50), and the noise power spectrum is shifted to lower frequencies with higher iterative strengths (51).

(17)

9

Figure 5. CT chest images with upper left: Filtered back projection, upper right: AIDR Enhanced (hybrid statistical algorithm), lower left: FIRST Lung (model-based algorithm), lower right: partition of the three algorithms filtered back projection, AIDR Enhanced and FIRST Lung.

Advanced Modeled Iterative Reconstruction, ADMIRE

Vendor advertised behavior: ADMIRE uses statistical weighting in the sinogram space to reduce spiral artefacts. Data is projected into image space. Here, a regularisation process with smoothing constrains is conducted together with a statistical model. This is done to separate noise from anatomical structures and reduce noise. Then the data is re-projected into pse do sinogram data sing a CT s stem model describing characteristics s ch as detector t pe and si e and fl ing focal spot (34). Five levels of iterative strength are available.

Peer reviewed literature characterization: Quantitative measurements are improved compared to filtered back projection and lower strengths (52, 53). Noise power spectrum is shifted to lower frequencies (54). ADMIRE 3 is shown in Figure 4.

(18)

10 Model-based algorithms

Vendor advertised behavior: Model-based iterative reconstruction algorithms apply forward projection to compare the artificial sinogram data with the measured sinogram data. Also, more complex models taking the acquisition process, image statistics and the scanner geometry and optics into account are applied. Both the forward process and optics modelling are time

consuming, and therefore model-based algorithms have longer reconstruction times compared to FBP, statistical and hybrid techniques.

Veo

Vendor advertised behavior: Veo was the first iterative reconstruction algorithm which modeled s stem optics and as considered a real model-based algorithm. Both the focal spot s and the detector element s ph sical e tent is modeled, a 3D voxel is imaged by a finite radiation beam originating from the focal spot and measured by the active detector area. Projection data is weighed so that noisy projections have lower influence than less noisy projections on the

reconstructed image. A priori knowledge of medical data is used to reduce noise in homogeneous areas while preserving edges around organ boundaries (55).

Peer reviewed literature characterization: Quantitative measurements are improved and the noise power spectrum shifts towards lower frequencies compared to both filtered back projection and ASiR (37, 56-58). Hounsfield units are shifted to lower values at low doses (59).

Forward projected model-based Iterative Reconstruction SoluTion, FIRST

Vendor advertised behavior: FIRST starts with an initial estimated image based on the measured projections. This image is forwarded into sinogram space where synthesized and measured projections are compared. Based on the difference between synthesized and measured

projections, the number of iterations to achieve the desired image quality is decided. Statistical and anatomical noise, scanner, optics and cone beam are modelled for an improved image. The synthesized data is back projected, and the process is repeated until the desired image quality is achieved (18).

Peer reviewed literature characterization: Quantitative measurements are improved compared to filtered back projection (46, 60). An image reconstructed with the FIRST Lung algorithm is shown at the lower left in Figure 5.

Artificial intelligence (AI) algorithms

In 2018, GE and Canon both introduced reconstruction algorithms utilising deep neural network, employed in artificial intelligence, to develop the algorithm. They were both cleared by the U.S.

Food & Drug Administration in 2019 (61, 62). The development of these algorithms consisted of three processes: 1) Network design, 2) Network training, 3) Validation period. The algorithm is now awaiting FDA approval for implementation in the clinic.

The neural network has an initial design. This consists of mathematical algorithms in layers incorporating knowledge about the imaging process, x-ray properties and system properties.

(19)

11

Neural networks can handle thousands of parameters, far more than what is possible for humans (63). Since the next step is training of the network, the initial network is not required to be exact and/or detailed. This is taken care of in the training process.

In the training process both input data and a gold standard are included in the network. The input data is low dose sinogram data, and the gold standard is a high dose, high quality data set of the same patient or phantom. The input data is fed to the network to obtain output data which is compared with the gold standard. From this comparison the network adjusts its parameters and a new data set can be inserted into the network. The process can be repeated several times until only small adjustments are needed, or the result is satisfactory (64, 65).

The validation process is important to ensure that the algorithm reproduces the information correctly. The algorithm must be tested for normal anatomy and for pathology, including rare pathology. When the algorithm is finalized and it is implemented on the CT scanner, the algorithm stays unchanged (65). A continuous evolvement of the algorithm could be very difficult to relate to for the radiologists, who often need older examinations for comparison. Properties of the artificial intelligence algorithms are shown in Table 2.

Table 2. Algorithms developed by using deep neural networks Algorithm –

Launched

Type Levels Training Reconstruction

time

TrueFidelity – 2018 (65)

Deep learning image reconstruction

Low, medium, high

High dose filtered back projection

Abdomen pelvic seconds

Advanced

Intelligent Clear-IQ Engine, AiCE – 2018 (64)

Deep learning reconstruction

Mild, standard, strong

High dose model- based iterative reconstruction

Five times faster than MBIR ->

upper abdomen 0.5 1 min

Since artificial intelligence algorithms can be trained with data sets possessing special properties, image quality can be guided in a desired direction. TrueFidelity is trained on high dose filtered back projection with the intention of maintaining the image texture and spatial resolution of filtered back projection images while reducing image noise. AiCE is trained on high dose model- based iterative reconstruction. Its aim is maintaining or reducing the image noise and improving the spatial resolution compared to the hybrid, statistical iterative reconstruction algorithm AIDR.

1.3 Dose descriptors

Radiation doses from CT are expressed by the CT dose index (CTDI) and the dose length product (DLP). The CT dose index is the measured radiation dose in a single axial rotation.

𝐶𝑇𝐷𝐼 𝑎 ∫ 𝐷 𝑧 𝑑𝑧₋ (4)

where D(z) is the radiation dose measured at position z. When measured with a pencil ionization chamber of 100 mm the CT dose index is denoted CTDI100. The CT dose index reported by the CT

(20)

12

scanners is measured in cylindrical phantoms with 16 cm (head) or 32 cm (body) diameters. Since the radiation beam is attenuated through the phantom, the dose in the middle is lower than the dose at the edges. For this reason, a weighted CT dose index is calculated. For helical scans a low pitch will result in higher dose, and a high pitch will result in lower dose. To take this into account the weighted CTDI is divided by the pitch, p, to get the CTDIvol.

𝐶𝑇𝐷𝐼 𝐶𝑇𝐷𝐼 ^𝑎 𝐶𝑇𝐷𝐼 ^𝑎 (5)

This is a measure of the axial average dose in the phantom and has the unit milli Gray [mGy]. To get an approximation of the radiation dose for the total scan, the CTDIvol is multiplied by the scan length.

𝐷𝐿𝑃 𝐶𝑇𝐷𝐼 ∗ 𝑠𝑐𝑎𝑛 𝑙𝑒𝑛𝑔𝑡ℎ (6)

The dose length product has the unit [mGy*cm]. Both CTDIvol and dose length products are reported after each scan. These values are related to a phantom and are not equivalent to the patient dose (66).

1.4 Clinical examinations

CT liver

Diseases in the liver can be caused by infections (hepatitis A and B are the most common), poisoning (alcohol, drugs, mushrooms), immunological causes or generalized disease (67).

Pathology of the liver parenchyma includes inflammation, fibrosis, cellular accumulations, cell death or regenerative changes (2). Contrast media is used to enhance the contrast of lesions and vessels compared to the liver parenchyma (Figure 6).

Figure 6. CT images of the liver. Without contrast (left) and with contrast (right).

The liver is a complex organ to image. The dual blood supply of the liver results in different enhancement at different times after contrast injection. The liver parenchyma is supplied 80%

from the portal vein and 20% from the hepatic artery, while tumors are mainly supplied from the hepatic artery (68). Hyperdense lesions (brighter than the liver parenchyma) will be best

(21)

13

visualised early, while hypodense lesions (darker than the liver parenchyma) will be best

visualised 60-90 seconds after injection. Lesion sizes can vary from 2-3 mm in diameter to several cm. Tissue contrast differences are often small, hence low contrast resolution is an important image quality parameter in liver imaging. To achieve a good low contrast resolution, it is

important to keep the image noise at a low level. Since lesions can be small, it is also important to reproduce their edges, making lesions more easily depictable.

Before surgery it is also important to evaluate vessel branching. This can have some of the same properties as lesions and low contrast to the parenchyma (though only hyperdense). Also, the smallest branches have small diameters; thus making the spatial resolution important.

The challenges with low contrast and small structures in the liver make this an interesting organ to optimise regarding image reconstruction. The benefits of low image noise may be lost with changes in image texture.

CT chest

Pulmonary diseases may be caused by infections, genetics and smoking. Examples are chronic obstructive pulmonary disease (inability to exhale normally), emphysema, bronchitis, cystic fibrosis, cancer, interstitial lung disease and pulmonary embolism (69). Pathology can be visible as for example fluid accumulation, parenchymal densifications or patterns, lesions and vessel occlusions.

In chest examinations, usually two sets of images are reconstructed: One set where high contrast resolution is emphasized (vessels and lesions against lung parenchyma) (Figure 7), and one where low contrast is emphasized (lung parenchyma and mediastinum). Most often, contrast media is used to enhance the contrast in vessels and to characterise lesions. These are seen in the lung parenchyma which to a large part consists of air, making the contrast difference between pathology and normal anatomy large. The vessels can be small, and to detect occlusions of small vessels it is important to have good spatial resolution. Lesions as small as 3-4 mm in diameter need to be detectable.

Chest is the organ besides abdomen/pelvic which is most frequently examined with CT (70). High contrast makes the image properties different from the liver adding value to the evaluation of iterative reconstruction algorithms.

Figure 7. CT image of chest.

(22)

14

1.5 Assessment of image quality

Image quality assessment is challenging since image quality consists of different properties which influence each other. In a clinical setting, pathology visualization is influenced by a combination of these image quality properties. Often image quality must be evaluated in patients without

pathology and then a correlation between image quality of normal structures and pathology is assumed. If quantitative measurements are used, each physical parameter is measured separately, and it is difficult to assess the combined effect.

Quantitative measurements are more consistent than evaluation based on human observers. This is useful when evaluating the quality of the equipment, and when new technology is introduced.

Quantitative measurement is the first step in describing the differences between new and existing technology (71, 72). However, the relationship between quantitative measurements and clinical image quality, or even diagnostic outcome, is not well known, and the correlation may be poor if the measurement strategy is not carefully chosen (71). Studies from planar x-ray show good correlation between visual grading analysis and quantitative measurements when the

quantitative measures are based on the Rose model (73, 74). This is not valid for reconstructed images in CT. The assumption that visualisation of normal anatomy correlates with the detection of lesions was also invalidated in planar x-ray (75) and were dependent on clinical task in

mammography (76).

When using image quality phantoms, it is even harder to obtain a clinically valid result. Phantoms are useful for benchmarking, standardisation and quality assurance testing. Furthermore,

phantoms can be useful for comparing and improving scan techniques without any radiation dose concerns. Image quality evaluation can be narrowed down to studying one specific image quality property at a time, for example low contrast resolution or the modulation transfer function.

However, the phantom, as well as the quantitative measure, must be chosen carefully so that the measured property is of relevance for the clinical task you are studying. Most often, several different quantitative parameters must be seen in combination to fully assess image quality. For example, the subjective reading of low contrast objects in a phantom study proved unsuitable for quality assurance testing on its own, since the radiation dose had to change with 50% before a change in detectability was seen (77). This is a high radiation dose change in the clinic, and in this situation the clinical image quality will be influenced before the quality assurance testing is finding changes. Doing testing which find errors or defects before clinical image quality is influenced is important.

Bayesian-based methods, the model observer like e.g. the Channelized Hotelling Observer or the Non-Prewhitening Matched Filter, are mathematical models trying to read images like human observers (78-80). Depending on the model design they may be useful tools for image quality assessment in phantoms in order to save time for the radiologist (81, 82). A combination of quantitative and human observer assessment may still be the most valuable.

Receiver operating characteristics

Receiver operating characteristics have been used in medical imaging to optimise the detection of pathology/lesion signals. A diagnostic test will have a discrimination threshold which

differentiates between healthy and diseased tissue. This threshold will result in some false positives and false negatives (Figure 8, left). Images or areas are scored based on how likely there

(23)

15

is a signal present (83, 84). From this the sensitivity (the probability of the test detecting the disease if the patient is sick)

𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 ^𝑁 ^𝑇𝑃

𝑁 𝑇𝑃 +𝑁 𝑎 𝑎 𝐹𝑁 (7)

and the specificity (the probability of the test showing a healthy person if the patient is healthy)

𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 ^𝑁 ^𝑎 ^𝑇𝑁

𝑁 𝑎 𝑇𝑁 +𝑁 𝑎 𝐹𝑃 (8)

can be calculated. The receiver operating curve can then be plotted as the true positive rate

Figure 8. Left: Probability density of a sick and healthy population with a test threshold resulting in true positive (TP); test shows that a sick person is sick, false positive (FP); test shows that a healthy person is sick, true negatives (TN); test shows that a healthy person is healthy, and false negatives (FN); test shows that a sick person is healthy. Right: Receiver operating characteristics (ROC) curves for each scenario. a) Curve A, b) Curve B, c) Curve C. The diagonal shows results that are acquired by chance.

(24)

16

which equals sensitivity against the false positive rate which equals 1-specificity (Figure 8, right).

The area under the curve is often used as a figure of merit. The larger the area under the curve, the better the image is for detecting the signal (74). Curve B and C have lower areas under the curve, and this is visualised to the left in Figure 8 by more overlapping curves between sick and healthy in b) and c). Curve B and C may have the same area under the curve, but they are differently skewed. Curve B has higher sensitivity at high specificity while curve C has higher sensitivity at low specificity. At the point they cross they have the same sensitivity and specificity.

By changing the decision threshold of the test to the left, you will get less true and false positives and more false negatives. This will result in lower sensitivity and higher specificity, hence moving down the curve. The opposite happens if you increase the decision threshold. Depending on the indication of the test and if one of the tests must be chosen, high sensitivity or high specificity may be favorable. High sensitivity is important when the disease is life-threatening. The failure to not diagnose a patient with a disease could lead to that person s death. High specificity may be of higher importance when most of the patients are healthy, e.g. in screening. Here, false positive results will generate additional unnecessary tests and anxiety for the patient.

Because it is the detection of pathology which is scored, this method is often regarded as the gold standard. Still, this method cannot be generalized to other organs or examinations. Another disadvantage is that you need to know the ground truth, that is, if the pathology or lesion detected actually is present or not. In patients this can often be difficult or could be a very time- consuming process. Therefore, anthropomorphic image quality phantoms can be useful.

Visual grading

In visual grading images are graded to reflect the perceived image quality of anatomical structures or image quality parameters. Predetermined quality criteria which fit the evaluation objective and images are used, for example EU guidelines on quality criteria for computed tomography (85).

Normal structures are scored assuming this correlates with the visibility of pathology.

Evaluation of criteria can be performed in different ways. Image sets can be scored individually.

This makes analysis of the score easier, but it is more difficult for the interpreter to maintain a consistent scale across the whole reading process, and different readers may interpret the scale differentl As a res lt the importance of distinct definitions of the score increases e g non- diagnostic s diagnostic Incorporating training sets in the beginning of readings is vital to decrease the differences in how the scale is interpreted.

A comparison between two image sets can be performed; either deciding which one is best or giving a score on how much better or worse one is compared to the other. This is effective when only one parameter is changed between image sets. When several parameters are changed, the method will result in many comparisons and statistical considerations become too challenging.

This can also introduce a scoring bias, depending on which image you compare to.

Visual grading is an easy way of grading images and the method is always feasible. One disadvantage with this method is the «beauty contest» aspect, which means that the prettiest image is favored without saying anything about the possibility of hitting correct diagnosis. Since the scale usually is an ordinal scale the analysis method can be visual grading characteristics if two techniques are compared (86), or visual grading regression if more techniques are compared (87, 88).

(25)

17 Quantitative measurements

Quantitative metrics commonly applied in CT imaging is image noise, signal to noise ratio, contrast to noise ratio, modulation transfer function and noise power spectrum.

Image noise

Image noise in CT is defined as fluctuation of CT numbers around a mean CT number. This can be expressed with the standard de iation

𝑁𝑜𝑖𝑠𝑒 𝜎 ^∑ ^{− ̅}

− (9)

where 𝑥 is the CT number of the pixel, 𝑥̅ is the mean CT number of the pixels, and n is the total number of pixels in the measured area (89). This property is the magnitude of image noise, but do not describe the frequency content of the noise. This is an important property, especially when examining objects with small contrast differences to the background.

Signal to noise ratio, SNR

The signal to noise ratio is defined in many ways. However, the most common definition uses the mean CT number in a region of interest, 𝑥̅, divided by the noise in the area, 𝜎 (79)

𝑆𝑁𝑅 ^̅ (10)

The metric is easy to apply, however, it does not take into account the size of the object and thus correlates poorly with human observers (74).

Contrast to noise ratio, CNR

The contrast to noise ratio can be defined as the difference in CT number between object, 𝑥̅ , and background, 𝑥̅ , divided by the noise, 𝜎, (sometimes referred to as signal difference to noise ratio)(79)

𝐶𝑁𝑅 ^{̅ − ̅} (11)

A different definition is provided by Thitaikumar et al. (90)

𝐶𝑁𝑅 ^{̅ − ̅}

+ , (12)

where σo is the noise in the object and σb is the noise in the background. Similar to the signal to noise ratio, contrast to noise ratio correlates poorly with human observers (74).

(26)

18 Low contrast resolution

Low contrast resolution is the ability to visualise objects with low contrast to the background. This is more difficult when objects get smaller, and the property is largely dependent on image noise (89). An often used method to determine the low contrast resolution is to use phantoms with objects of different contrast and sizes, and to find visually the smallest object at each contrast level.

Noise power spectrum, NPS

The 2D noise power spectrum is defined by Verdun et al. (79) as 𝑁𝑃𝑆 _𝐷 𝑓_,𝑓 ^{∆ ∆}

𝑁 ∑^𝑁₌ |𝐹𝑇 _𝐷 𝑅𝑂𝐼 𝑥, 𝑦 𝑅𝑂𝐼 | (13) were x and y are pixel sizes and nx and ny are the number of pixels in each dimension of the region of interest (ROI), NROI is the number of regions of interests used in the averaging process, FT2D is the two dimensional Fourier transform, 𝑅𝑂𝐼 𝑥, 𝑦 is the CT numbers in the ith ROI, and

𝑅𝑂𝐼 is the mean CT number in the ROI. The noise power spectrum shows the spatial frequency content of image noise. This is important for the texture in the image. Figure 9 shows noise power spectra with low and high frequency content. A high content of low frequency noise will give a more blotchy appearance (Figure, 9, left) compared to images with high frequency noise (Figure 9, right).

Modulation transfer function, MTF, and spatial resolution

The modulation transfer function is defined in a linear, shift-invariant system, in terms of the 3D system point spread function, p(x,y,z), as

𝑀𝑇𝐹 𝑢, 𝑣, 𝑤 |𝐹 𝑝 𝑥, 𝑦, 𝑧 | (14)

where u, v and w are spatial frequency variables corresponding to the spatial variables x, y and z (91, 92). This meas re describes the s stem s abilit to present contrast at different freq encies hence ho the imaging s stem transfers an objects spatial freq enc content into the image domain. It is used to describe the spatial resolution of the system. An example is shown in Figure 10.

Since iterative reconstruction algorithms make systems non-linear, the modulation transfer function will vary with contrast and radiation dose (79). Hence a task-based modulation transfer function was introduced (93). This takes the specific task into account, with a specific contrast and noise level.

1.6 Linear versus non-linear reconstruction algorithms

In images reconstructed with filtered back projection and standard reconstruction kernels which do not enhance edges or any areas of the Hounsfield numbers, the image noise is inversely proportional to the square root of the radiation dose (79). This relation is independent of dose in

(27)

19

Figure 9. Liver phantom reconstructed with Veo (low frequency content) to the left and with filtered back projection (high frequency content) to the right. Image noise measured in the liver was 10 HU (left) and 14 HU (right). The lower plot shows corresponding normalised noise power spectra, the left curve illustrating the texture in the upper left image and the right curve illustrating the texture in the upper right image.

Figure 10. a) Two points to image, b) Point spread function of the two points, c) Modulation transfer function of the two points.

(28)

20

the clinical dose range, and in this thesis, this is referred to as linear reconstruction. Iterative reconstruction algorithms may be adaptive, hence reducing image noise more at certain dose levels compared to other dose levels, preferably at low doses. This will also affect signal to noise ratio and contrast to noise ratio (94, 95), and the changes in properties will be dependent on slice thickness (96). Spatial resolution and image texture can also change with radiation dose (or image noise) and contrast (93, 97-99). Hence, iterative reconstruction is referred to as non-linear

algorithms.

Quantitative measurements like the ones previously described, were more frequently used as figures of merit of image quality with filtered back projection. The modulation transfer function requires a shift-invariant system meaning that resolution properties should not change across the image. With iterative reconstruction algorithms the non-linear and non-stationary properties are stronger than with filtered back projection, making Fourier analysis like modulation transfer function and noise power spectrum more challenging (100). It becomes of greater importance to perform measurements at the same position and with the same contrast when making

comparisons between images. Change in image texture associated with iterative reconstruction may result in almost noise free images without necessarily increasing lesion conspicuity. It becomes more important to describe noise frequencies, but this should be done carefully using small ROIs and at the same position (79). Additional evaluation of clinical structures may be necessary.

1.7 Image quality phantoms

The CT vendors have developed their own image quality phantoms, which are designed for use during maintenance of CT scanners. These phantoms are circular, with diameters usually between 16-20 cm, having a homogenous insert made of water or plastic material. They can be used for measurement of image noise differences and ring or streak artefacts. Some of the phantoms also have modules for testing Hounsfield units for different materials, modulation transfer function, and low contrast resolution.

Technical phantoms, e.g. the Catphan image quality phantoms (The Phantom Laboratory, Salem, NY, USA) are used in quality assurance testing (101), and to some extent in clinical testing, e.g. of CT head, metal artefact reduction or in small patients (102-106). These phantoms are specially designed for technical assessment of different quantitative image quality parameters according to international guidelines and recommendations on quality assurance testing of CT scanners. Most of the vendors technical specifications are referring to the Catphan phantoms, and physicists worldwide use these for quality assurance testing of CT scanners. Still, parameters used for quality assurance tests may be very different from parameters applied for a clinical examination, both with respect to size, shape, density, structures and lack of anatomical noise and movement.

In addition, most of the technical phantoms in use today were designed when filtered back projection was the main method of reconstruction, not considering the non-linearities of iterative reconstruction. Therefore, more patient-like phantoms are needed for optimisation of image quality in a clinical setting.

Anthropomorphic phantoms are more patient-like with respect to shape and size, including anatomical structures inside the phantom. These are more suitable for testing clinical protocols.

The Lungman (Kyoto Kagaku Co, Kyoto, Japan) ), the ATCM phantom and the multipurpose liver phantom (The Phantom Laboratory, Salem, NY, USA) and QRM liver phantom (QRM Quality

(29)

21

Assurance in Radiology and Medicine GmbH, Moehrendorf, Germany) are all examples of this type of phantom (101).

With an increase in CT examinations that are non-linear (both iterative and deep learning image reconstruction) the need for phantoms combining quantitative measurements and human observer assessment is emerging. In collaboration with The Phantom Laboratory, we have developed an upper abdomen phantom with inserts where both ROC readings and quantitative measurements can be assessed in the same scan (107-109).

(30)

22

(31)

23

2 Aims of the thesis

The aim of this thesis was to evaluate new reconstruction technology in CT in chest and liver examinations with respect to image quality parameters and radiation dose dependency.

Specific aims:

1) To compare performance of new CT image reconstruction techniques with the performance of the established filtered back projection technique.

2) To evaluate the outcome of different image quality evaluation methods used on images reconstructed with iterative reconstruction, including the correlation between

quantitative measurements, visual grading analysis and receiver operating characteristics.

(32)

24

(33)

25

3 Material and method 3.1 Phantoms

Anthropomorphic image quality phantoms were used to test iterative reconstruction algorithms.

For abdominal examinations a liver phantom c stom made b St Bartholome s Hospital Clinical Physics Group, London, UK) was used. The phantom is shown in Figure 11.

Figure 11. Liver phantom (St. Bartholomew’s Hospital, Clinical Physics Group, London, UK) (upper row), and chest phantom (Lungman, Kyoto Kagaku Co, Kyoto, Japan) (lower rows).

(34)

26

The liver phantom used in paper 1 and 3 allowed variation of both pattern and densities of the lesions in the liver. Due to the fact that comparisons were conducted across vendors, it was decided to use water, as the Hounsfield units of water should be 0 ±4 HU for all vendors according to international guidelines (110). Hounsfield units for other densities may vary depending on total filtration and reconstruction kernels used. The lesions made it possible to use receiver operating characteristic analysis of the images. The phantom also had anatomical structures resembling a kidney, pancreas, spine and a liver, and these anatomical structures were used for visual grading analysis in addition to receiver operating characteristics in paper 3.

For chest examinations (paper 2 and 4) the Lungman phantom was used (Kyoto Kagaku Co, Kyoto, Japan), Figure 10. The phantom comes with a set of 15 lesions, all with three different Hounsfield units (-800, -630, +100) and with five different diameters (3-12 mm) that can be inserted in the thoracic area in between the pulmonary vessels. These lesions are easy to spot, even at radiation doses as low as 1 mGy. Thus, to use receiver operating characteristics in order to depict changes in radiation dose with this phantom, is not constructive. Instead, visual grading analysis of anatomical structures and general image quality properties was applied, and the diaphragm as a homogenous area was used for calculating the noise power spectrum.

The advantage with phantoms is that size and densities are the same for each scan, facilitating a systematic review and comparison of algorithms for different scanners at different radiation dose levels. Since radiation doses are associated with an increase in risk of developing cancer, we cannot scan a patient numerous times. Furthermore, the same patient cannot be scanned on several different scanners in different locations. For one thing, variation in patient size and anatomy is large. In addition, differences in dose modulation techniques on different CT models will influence image quality. Together with the properties of reconstruction algorithms, it is difficult to separate the different effects on image quality. These uncertainties are avoided when using an image quality phantom.

3.2 Analysis methods

Receiver operating characteristics were used when analysing images taken of the liver phantom (paper 1 and 3). A four-point scale was used with scores 1: No lesion, 2: Possibly no lesion, 3:

Possibly lesion and 4: Lesion. A total of 32 sectors were scored in each image with lesions in 16 of the sectors. Positions of the lesions were changed to avoid memory bias. True positive fraction was plotted against false positive fraction and the area under the curve was calculated using a nonparametric methodology using the statistical software Analyse-it (Analyse-it Software, Ltd., Leeds, United Kingdom). A comparison of different curves was conducted using the Delong- Delong method (111). This method was used to show differences in lesion conspicuity with different reconstruction algorithms.

In paper 3 and 4, visual grading was used to score image quality. A five-point scale was applied, and anatomical structures and image quality parameters were scored. Data was analysed using ordinal logistic regression with the command meologit in the statistical software STATA

(StataCorp LLC, College Station, Texas, USA). This command allowed for fixed and random variables. Observer was treated as a random variable and reconstruction algorithm as a fixed variable. This method was used to show differences in the presentation of anatomical structures and image quality parameters from the algorithms.

(35)

27

Quantitative measurements used were image noise, signal to noise ratio, contrast to noise ratio and normalized noise power spectra. These were used to understand differences in images reconstructed with different techniques and to support and explain the clinical evaluations of image quality.

(36)

28

(37)

29

4 Results

4.1 Paper 1: Comparing five different iterative reconstruction algorithms for computed tomography in an ROC study

Veo improved lesion detection in the liver at all dose levels compared to filtered back projection.

SAFIRE 3 improved lesion detection at the lowest dose level, 5 mGy, compared to filtered back projection. ASiR, iDose and AIDR 3D did not improve lesion detection significantly compared to filtered back projection. Significant differences between algorithm strengths were not seen.

Veo resulted in better lesion detection in the liver than ASiR and AIDR 3D at 5 mGy, better than all the algorithms except the highest strength of SAFIRE at 10 mGy, and better than ASiR and iDose at 15 mGy. SAFIRE showed better lesion detection than ASiR, iDose and AIDR at 5 mGy, better than iDose and AIDR 3D at 10 mGy and better than iDose at 15 mGy. iDose showed better lesion detection than AIDR 3D at 5 mGy. AIDR 3D resulted in better lesion detection than ASiR at 15 mGy.

All the algorithms improved image noise and contrast to noise ratio, but only Veo and SAFIRE showed an improved detection rate, e.g., AIDR 3D strong showed an image noise reduction of 64% compared to filtered back projection and contrast to noise ratio increased from 0.8 to 2.4, but detection rate was the same.

In conclusion, lesion conspicuity varied with different iterative algorithms, and the performance depended on dose level.

4.2 Paper 2: Image quality in oncologic chest computerized tomography with iterative reconstruction. A phantom study

ASiR increased contrast to noise ratio at all dose levels (6-65%) except for two lesions at 6.7 mGy and for one lesion at 0.6 mGy with ASiR 30%. A reduction in contrast to noise ratio was observed (4-20% and 34% respectively). Veo increased contrast to noise ratio at all dose levels, but the increase was higher at lower dose levels. ASiR improved signal to noise ratio compared to filtered back projection at the three highest dose levels, but no big difference was seen at 0.6 mGy. Veo improved signal to noise ratio at all dose levels compared to filtered back projection. Hounsfield units and range were similar for filtered back projection and ASiR. They did not differ much between dose levels except for the lowest dose level where the range increased compared to higher dose levels. This may be due to photon starvation and cupping artefacts. In the muscle, Veo reduced the range of Hounsfield units at the low dose level compared to filtered back projection and ASiR, indicating a reduction in artefacts from photon starvation and cupping artefacts. Veo shifted Hounsfield units towards negative values.

In conclusion, quantitative measurements were improved with ASiR and Veo compared to filtered back projection and with Veo compared to ASiR. Veo reduced cupping artefacts but shifted Hounsfield units towards lower values.

Iterative reconstruction in CT imaging; Image quality and radiation doses