Homogeneous Canine Chest Phantom Construction: A Tool for Image Quality Optimization

Digital radiographic imaging is increasing in veterinary practice. The use of radiation demands responsibility to maintain high image quality. Low doses are necessary because workers are requested to restrain the animal. Optimizing digital systems is necessary to avoid unnecessary exposure, causing the phenomenon known as dose creep. Homogeneous phantoms are widely used to optimize image quality and dose. We developed an automatic computational methodology to classify and quantify tissues (i.e., lung tissue, adipose tissue, muscle tissue, and bone) in canine chest computed tomography exams. The thickness of each tissue was converted to simulator materials (i.e., Lucite, aluminum, and air). Dogs were separated into groups of 20 animals each according to weight. Mean weights were 6.5 ± 2.0 kg, 15.0 ± 5.0 kg, 32.0 ± 5.5 kg, and 50.0 ± 12.0 kg, for the small, medium, large, and giant groups, respectively. The one-way analysis of variance revealed significant differences in all simulator material thicknesses (p < 0.05) quantified between groups. As a result, four phantoms were constructed for dorsoventral and lateral views. In conclusion, the present methodology allows the development of phantoms of the canine chest and possibly other body regions and/or animals. The proposed phantom is a practical tool that may be employed in future work to optimize veterinary X-ray procedures.


Introduction
The use of radiation in veterinary medicine demands responsibility to maintain rigid radiation safety standards and practice. Workers are often requested to restrain animals during the procedure to avoid movement artifacts and maintain image quality [1,2]. Therefore, the exposure of workers have to be minimized considering the radiation protection principles because of their presence in the examination room throughout the procedure [3][4][5]. In such situations, scattered radiation represents the main source of radiation that is received by operators [1].
With the increasing use of digital radiographic imaging in medical and veterinary practice, a phenomenon known as dose creep has been reported [6,7]. Dose creep is the gradual increase exposure factors and dose over time for a given radiographic anatomical projection. The phenomenon occurs due to the ability of the wider dynamic range of the system and computer processing of the acquired image to correct inappropriate exposure parameters [7][8][9].Therefore, optimize digital systems to maximize image quality and minimize personnel exposure is important [3].
Optimization procedures have been widely exploited for humans in the literature [1,[10][11][12][13][14] Homogeneous phantoms are the simplest tool that is used for optimization purposes [15,16]. Homogeneous phantoms are extensively employed to assess objective metrics associated with image quality, such as the signal difference-to-noise ratio (SdNR) [12,14]. A higher SdNR provides superior image quality compared with a lower SdNR [12,14].
These phantoms are constructed of tissue-equivalent materials with tridimensional structures to simulate the absorption and scattering of X-ray beams in the body. Homogeneous phantoms are generally constructed of polymethyl methacrylate (Lucite) and aluminum through the classification and quantification of tissues from computed tomography (CT) exams [15].
However, the development of homogeneous phantoms for veterinary patients has remained unexplored. Therefore, the aim of this work was develop homogeneous canine thorax phantoms for the dorsoventral and lateral views for different sizes of dogs. These phantoms may be applied for optimization processes and to improve image quality, reduce personnel exposure minimizing the occurrence of dose creep.

Database
The present study was developed with approval from the Ethics Committee on the Use of Animals of the authors' institutions and national review panels under protocol no. 90/2014. The database consisted of 356 retrospective chest CT exams of dogs, which were veterinary patients from the Diagnostic Imaging Section of the Veterinary Hospital, University of Veterinary Medicine and Animal Science (HV/FMVZ-Unesp Botucatu), between January 2007 and May 2015. Veterinary patients with significant thoracic disease, including primary pulmonary masses, pneumothorax, or pleural effusion, were excluded. Veterinary patients with superimposed nodules or a high number of pulmonary nodules to be counted reliably were also excluded. Therefore, 276 patients were excluded as described, leaving 80 for following analyses.
The veterinary patients were separated into groups of 20 dogs each according to weight: small (S) group ( 9.

Automatic computational algorithm
We used an established methodology [10,15,18] to develop an automatic computational algorithm that is able to classify and quantify biological tissues of the canine chest based on retrospective CT exams. The slices from CT exam were used to obtain the amount of each biological tissue present in the studied region, which was converted in simulator materials to construct the homogeneous phantoms. For each canine group, we developed two types of phantoms (dorsoventral chest phantom and lateral chest phantom) to simulate two common projections that are used in clinical practice. The construction of the phantoms is described below.
Histogram of CT exams. Each CT slide was automatically aligned to correct possible dog positioning errors. Based on differences in the attenuation of each tissue in the CT slice ( Fig  1A), represented by CT number (Hounsfield Units [HU]), the algorithm classified each voxel as belonging to a given tissue. This classification was based on the histogram of CT number. Fig 1B illustrates the histogram of image shown in Fig 1A that presents four well-separated characteristic peaks of different tissues: lung tissue, adipose tissue, muscle tissue, and bone.
Classification and quantification of chest tissues by Gaussian curves. To automatically classify each voxel from the histogram, the algorithm used Gaussian distribution curves as the base. These curves represent the degree of accuracy of a voxel that belongs to a given tissue (adipose tissue, muscle tissue, lung tissue, and bone) according to CT number. An experienced veterinary radiologist manually measured the mean and standard deviation of the CT number for each tissue. The measured values were used to create Gaussian distribution curves.  Fig 1B divided by the curves that were created: -600 HU indicates lung tissue, -100 HU indicates adipose tissue, 100 HU indicates muscle tissue, and > 300 HU indicates bone. Trabecular and cortical bones presented significant differences in X-ray attenuation and thus were divided into two Gaussian curves with peaks at 300 HU and 400 HU, respectively. Therefore, each voxel that belonged to a histogram was classified as a specific tissue according to the Gaussian distribution curve.
To estimate the amount of tissue (i) in the exam, the number of voxels (N v ) that belonged to specific tissues (i) was used. N v (i) was calculated by multiplying the examination histogram by the corresponding Gaussian distribution curve. N v (i) was used to obtain the mean tissue thickness according to the following equation: where T(i) is the mean thickness of tissue (mm), i is the tissue type (lung, adipose, muscle, or bone), A p is the pixel area (mm 2 ), S is the number of slices, and D is the mean diameter of the chest CT exam (mm). The multiplication of N v with A p results in the total tissue (i) area (mm 2 ). The mean tissue area (mm 2 ) for each slice results from the total tissue area divided by S. To obtain the thickness of tissue (T(i)) in the X-ray beam direction, the mean area is divided by the mean perpendicular diameter, i.e. to develop a dorsoventral type phantom, the perpendicular diameter used in the equation was the mean lateral diameter. Likewise, to develop a lateral type phantom, the dorsoventral diameter was employed.
Finally, the mean thicknesses of the biological tissues were converted into the corresponding thicknesses of the simulator material plates. Lungs were simulated by air. Soft tissue (muscle tissue and adipose tissue) was simulated by Lucite. Bone (trabecular and cortical) was simulated by aluminum. This process was based on literature [19] using quantification values [20] for the composition of biological tissues.

Development of homogeneous canine chest phantoms
The dorsoventral and lateral homogeneous canine chest phantoms were constructed using the achieved thicknesses of simulator material and are shown in Fig 3. The dorsoventral homogeneous canine chest phantom (Fig 3A) was constructed of aluminum and Lucite plates. The first Lucite pair (on the top) sandwiched an aluminum plate. The Lucite thickness was distributed into four equal slabs. The second Lucite pair (on the bottom) contained a slightly wider aluminum slab inside. Between the upper and lower pairs of plates, spacers were inserted that represented lung tissue as an air gap. The architecture of the canine chest phantom is similar to an equivalent patient phantom [16,21] that is used for optimization purposes in human exams.
The architecture of the lateral homogeneous canine chest phantom was based on literature [22]. As shown in Fig 3B, the phantom was separated into three parts. Each part had an aluminum plate sandwiched by two Lucite plates. The aluminum that was placed in the intermediate pair accounted for the amount of bone from the vertebral column and extern. Air gaps were inserted between the pairs. The dimensions (length and width) of the phantoms were established as the mean dorsoventral and lateral-lateral canine size. The thicknesses and dimensions of the plates were separately calculated for each canine group.

Statistical analysis
The thicknesses of each simulator material was compared between different dog groups to assess whether the phantoms follow the same division of canine groups based on weight.
The statistical analysis of the Gaussian distribution data was performed using one-way repeated-measures analysis of variance (ANOVA) followed by Student's t-test for comparisons between canine groups for each simulator material. Values of p < 0.05 were considered statistically significant. This statistical analysis was performed for both phantom types (dorsoventral and lateral). The results are presented as mean and standard deviation.

Results
Tables 1 and 2 show the thicknesses for both phantoms (dorsoventral and lateral) according to canine group. Each canine group was characterized by the average thickness of biological tissue (Ti), thickness of the simulator material (TS), coefficient of variation (defined as the standard    deviation of the obtained examination thickness divided by the mean value), and mean dimensions (length x width) of the phantom. For both phantom types (dorsoventral and lateral), the one-way ANOVA revealed significant differences in Lucite, aluminum, and air thicknesses between groups (p < 0.05). Student's t-test revealed significant differences between groups (p < 0.05) for all materials, with the exception of aluminum and air between the L and G groups (Figs 4 and 5).

Discussion and Conclusion
Optimizing digital radiographic technique settings is important for animal imaging to maximize image quality while minimizing personnel radiation exposure [3] according to As Low As Reasonably Achievable (ALARA) principles.
In the present study, a robust algorithm widely used in the literature [10,15,18] was adapted for veterinary patients and implemented to classify and quantify the amount of tissue in the canine chest. As a result of the quantification, dorsoventral and lateral homogeneous canine phantoms were constructed according to canine group. Tables 1 and 2 show significant coefficient of variation. Such variations within the same canine group may be attributable to differences in nutritional status, breed, and age. Furthermore, deviations in the mean air thickness can result from changes in lung volume during inspiration and expiration.
The statistical analysis for both phantom views revealed significant differences between canine groups for all simulator materials. Student's t-test revealed significant differences for all of the Lucite analyses. For aluminum and air, all of the differences were significant, with the exception of the L and G groups. Such results may be attributable to the same factors that caused significant coefficient of variation. Differences revealed by statistical analysis and variations in the dimensions of the phantoms (length and width) were the main reasons for constructing four different phantoms for each phantom view.
Many studies have been performed to optimize X-ray exams using homogeneous phantoms for humans [10,12,14,23]. However, optimization is still neglected in veterinary medicine. Constructing a homogeneous canine phantom is important to adhere to ALARA principles and minimize dose creep. These phantoms may be used to optimize image quality and dose using such metrics as the signal difference-to-noise ratio, figure of merit [12,14], modulation transfer function, and effective detective quantum efficiency [10,24].
In conclusion, we developed eight phantoms of the canine chest using established methodology. Although we used a canine chest model, this methodology may be implemented for other anatomical structures and/or animals. Each of these phantoms were constructed of readily available materials [10].
The homogeneous canine chest phantom that was developed herein is a practical tool that may be employed in future work to optimize veterinary X-ray procedures and avoid the dose creep phenomenon.