Computed tomography (CT) scanners are expensive imaging devices, often out of reach for small research groups. Designing and building a CT scanner from modular components is possible, and this article demonstrates that realization of a CT scanner from components is surprisingly easy. However, the high costs of a modular X-ray source and detector limit the overall cost savings. In this article, the possibility of building a CT scanner with available surplus X-ray parts is discussed, and a practical device is described that incurred costs of less than $16,000. The image quality of this device is comparable with commercial devices. The disadvantage is that design constraints imposed by the available components lead to slow scan speeds and a resolution of 0.5 mm. Despite these limitations, a device such as this is attractive for imaging studies in the biological and biomedical sciences, as well as for advancing CT technology itself.
Computed tomography (CT) is an X-ray–based cross-sectional imaging modality that has a wide variety of applications in the biomedical and engineering sciences. The fundamentals of CT image formation have been described in BI&T.1,2 Unfortunately, the link from theory to a practical realization of a CT scanner often is lacking. Commercial CT scanners, ranging from more than $100,000 for small-animal scanners to several million dollars for whole-body scanners, often are beyond the reach of small research departments or units. Sometimes, on-site CT scanners are available for an hourly fee, which can easily reach several hundred dollars per hour. These costs do not encourage experimentation.
A visit to the surplus store may reveal useful X-ray equipment that can be used in place of new equipment to further cut costs.
One alternative is to build a CT scanner from parts, and one objective of this article is to demonstrate that creating a functional three-dimensional (3D) CT scanner from components is surprisingly easy. However, the cost for X-ray components remains relatively high; an X-ray generator costs nearly $50,000, and a suitable detector is approximately $10,000. By using all-new components, total costs for the complete CT scanner can potentially remain below $100,000, but further reduction of the overall costs would be difficult to achieve.
Many institutions, including hospitals and universities, manage surplus equipment. Often, surplus equipment can be reacquired by departments within the same institution, and the subject of surplus equipment has been briefly covered in BI&T.3 Thus, a visit to the surplus store may reveal useful X-ray equipment that can be used in place of new equipment to further cut costs. Depending on the available surplus equipment (availability of an X-ray generator is the minimum prerequisite), the design of the CT scanner needs to be adapted, and the final features of the scanner depend on said equipment. Therefore, the second objective of this article is to describe a specific design based on a surplus dual-energy X-ray absorptiometry (DEXA) scanner and its X-ray generator/detector system.
The basis for the CT scanner described in this article was a Hologic QDR1500 DEXA scanner (Hologic, Bedford, MA), which was considered inoperable because its controlling computer had failed and the company no longer offered to service the device.
Materials and Methods
From the Hologic DEXA scanner, the X-ray tube assembly with the associated high-voltage board and the detector were extracted. These were the only components from the Hologic QDR1500 that were used in this project. All other required components, including the mechanical motion stages, control electronics, radiation shielding, and image reconstruction computer, were acquired elsewhere.
The Hologic tube assembly is a lead-shielded and oil-filled tank that contains the X-ray tube itself and a high-voltage transformer. The tank assembly was used without further modification. A separate control board that provides the drive current for the high-voltage transformer from a 120-V AC utility line also was used without modification. The complete tube assembly in this form provides alternating X-ray pulses of 4 ms duration at 70 and 140 kVp, synchronized with the 60-Hz current of the AC mains. Tube current was specified as 1 mA, and feedback control of the tube current was provided on the control board. Moreover, the tube assembly included a Geneva wheel with multiple apertures that allowed selection of collimation apertures from 0.6 to 3.0 mm.
The detector module of the Hologic QDR1500 was a point detector, consisting of a CdWO4 scintillator crystal of approximately 2.5 cm by 2.5 cm side length that was coupled into a Hamamatsu R647-25 photo-multiplier tube (Hamamatsu, Hamamatsu City, Japan). The scintillation crystal, photo-multiplier, and high-voltage driver circuit, which provides up to 1,000 V for the photo-multiplier cascade, are enclosed in a compact, light-proof box, and this module also was used without further modification.
The tube tank was rotated by 90° to allow the beam to be emitted sideways. The tank and the detector box were mounted on opposing ends of an optical table. In addition, a 5-mm brass plate was mounted on a motor-driven x-y stage in front of the detector, and two selectable apertures of 0.5 and 2.5 mm, drilled into the brass plate, provided detector-side collimation and suppression of scattered X-rays.
Because of the weight of the X-ray assembly (90 kg), a decision was made to place the sample on a motorized stage and to keep the tube-detector system stationary. By convention, the axis of rotation is defined as the z-axis (axial direction). When the X-ray beam is considered to move along the y-axis, sample motion along the x-axis is necessary to obtain projection scans. Two linear slides with 30 cm travel (item no. 37-368; Edmund Optics, Barrington, NJ) allowed motion in the x direction. A Thorlabs optical breadboard (12.7 mm thick aluminum, item no. MB2530/M; Thorlabs, Newton, NJ) was placed on the linear slides to support a linear actuator (LS-100-06A; Anaheim Automation, Anaheim, CA). The latter was mounted upright for sample placement along the z direction. A motorized rotary stage (item no. 8730; Sherline, Vista, CA) was attached to the z actuator by means of a custom L bracket. The Sherline rotary stage comes with various means for sample attachment. Last, a linear actuator (23A104C with 12-in leadscrew; Anaheim Automation) was used to actuate the entire assembly in the x direction. Thus, three step motors are available to move the sample in the x and z directions and to rotate the sample. Figure 1 shows a sketch of the mechanical arrangement of these components.
Because the smallest collimation aperture is 0.6 mm, a point-spread function of no better than 0.6 mm can be expected. Therefore, a voxel size of 0.5 mm was chosen, which requires the x motor to perform exactly 16 half-steps between two zero crossings of the AC line. To minimize sample deformation and avoid step motor slippage, the motor controller was designed to accelerate the motor to its final speed of 1,920 half-steps per second in a linear, trapezoidal profile, and to reach phase synchronization with the AC line cycle at full speed. Acceleration was limited to 0.056 m/s2. Motion of the θ and z motors is not time critical and takes place independent of the 60-Hz cycle.
The computing power was provided in anticipation of future needs, and a much more modest PC would have been sufficient for instrument control and basic image reconstruction.
Design criteria for the control circuitry were dictated by the dependency of the tube on the 60-Hz AC cycle. More specifically, the custom control circuitry was required to provide several crucial timing signals: a zero-crossing signal that is active 1 ms before to 1 ms after the AC signal changes its sign and a control signal that is active during the central 4 ms of each AC half-cycle. Only during this 4-ms period, the X-ray tube actually emits radiation. Moreover, during this 4-ms period, the voltage is read from the photomultiplier tube.
At the end of the 4-ms exposure period, intensity data are collected and temporarily stored in the controller memory. After each profile acquisition, line scan data are sent to the host computer. Next, the sample is rotated by a predetermined angle (Δθ). Projection scans take place in alternating directions, and every other scan is reversed before storage. Acquisition, timing control, and motor control are performed by three microcontrollers (Figure 2), which in turn are controlled by a host computer that is responsible for high-level control, such as z positioning, initiating a scan, and collecting projection data.
During a tube warm-up phase of 5 s, the tube voltage is reduced to 40 kV while the cathode heating filament reaches its design temperature. Without the warm-up phase, the tube current control would overshoot the filament temperature and thus reduce the life span of the X-ray tube. To monitor tube aging and tube usage, a running time meter was added such that it displays the net time during which the tube is energized.
A standard personal computer (PC) was used for high-level data collection, instrument control, and image reconstruction. A fairly powerful computer was assembled from modules (AMD Phenom II 3.3 GHz six-core processor with 8 GB memory and a GPU-computing capable GeForce GT-430 graphics card on an ASUS M5A97 motherboard, 500 GB hard disk, and suitable PC power supply). The computing power was provided in anticipation of future needs, and a much more modest PC would have been sufficient for instrument control and basic image reconstruction. The reconstruction PC, the instrument power supply, Hologic high-voltage board, and control electronics each were provided with a rackmount case and joined in a 19-in industry standard rack. The rack also was used to mount the keyboard and computer display.
Software for instrument control and data collection was custom written in the C language and run on the PC under the Linux operating system. This software communicates with the low-level control components through the RS-232 serial interface with a simple command-data-checksum protocol. A graphical user interface for machine control was created, and its elements are shown in Figure 3.
Projection data are collected on the host computer as 16-bit intensity values for each scan line and each rotation angle (sinogram). Because each projection provides intensity values at 70 and 140 kVp, two independent sinograms are stored per slice and reconstructed independently. Intensity is averaged over the leftmost and rightmost 5 mm of each scan and used as the unattenuated reference intensity value I0. For each scan, absorbance (A) is computed from each discrete intensity value I as A = −ln(I/I0.4 A fast Fourier–based preview is available, but the actual reconstruction uses either the parallel beam–filtered backprojection algorithm5 or alternatively arithmetic reconstruction techniques.6 The reconstructed images represent apparent X-ray attenuation and need to be converted to Hounsfield units in a separate step.
For 3D scans, the sample is scanned over a total rotation of 180° or 360° to obtain one cross-sectional slice. After two-dimensional (2D) slice acquisition is finished, the sample is repositioned in the z direction for the next slice. In this fashion, volumetric scans are assembled from 2D cross sections. A variable slice-to-slice distance (Δz) is possible, and with Δz = 0.5 mm, voxels are isotropic.
Software for instrument control and data collection was custom written in the C language and run on the PC under the Linux operating system.
The original Hologic DEXA scanner was allowed to operate without additional shielding, and radiation levels, even when scattering inside the patient is considered, remained below the legal limit of 50 μSv per hour and 20 μSv at a distance of 25 cm. Nonetheless, a complete enclosure was built for the CT scanner, which provides 3 mm lead shielding at both side walls and at the detector face wall, and 1.5 mm lead shielding at the top and back, where the X-ray tank provides additional shielding. Lead-lined plywood was ordered to specifications from Mayco Industries (Birmingham, AL) and mounted on a custom steel frame. A sliding door with overlapping lead-lined edges was provided to allow sample access. An interlock system that was combined with an emergency shutoff button and a key lock allows the unit to operate only when the door is closed. The interlock system is completely hardware based, and interruption of the current at the interlock switch, emergency stop switch, or keylock directly shuts off the power supply of the high-voltage board. Thus, the interlock system becomes independent from the microcontrollers and their software.
The CT scanner was built in three distinct design stages. The first stage was the development of the timing controller (Figure 2) to which all timing-critical components synchronize. The timing controller was the prerequisite to test functionality of the X-ray tube and the X-ray detector. With proper functioning of the X-ray subsystem ascertained, the sample motion subsystem was built and finally the shielding enclosure added. Figure 4 shows the final CT scanner. Total costs were approximately $16,000, which included compensation for student workers (about $5,000). The lead-shielded enclosure ($3,000), optical table ($2,700), and sample motion components ($3,500) were major contributors to the overall cost. Time commitment by the author is not included in this calculation.
After the device was operational, it was immediately approved by the Radiation Safety Division of the University of Georgia, because radiation levels outside the enclosure were below background radiation levels and therefore not detectable. No difference in radiation levels was found near the final CT scanner when the X-ray beam was turned on or off. However, Georgia state law requires that the device becomes inoperable when the X-ray warning lamp fails. For this reason alone, the existing X-ray warning lamp that was integrated with the X-ray driver board was replaced by a lamp, for which the current could be measured by the main controller. With this feature, the main controller briefly tests whether the lamp is functional before enabling the X-rays and aborts the operation if no current through the lamp can be detected.
Alignment of the mechanical motion system is of paramount importance. The two most critical alignment tasks were the angular alignment of the z-axis and the adjustment of the scan center. For alignment of the z-axis, a steel rod was screwed into the center hole of the rotary stage, and scans taken at different vertical (z) positions. The angle of the rotary stage with respect to the y-axis was adjusted such that the edges of the steel rod in the projection stayed within the same window (±1 pixel) along the full height of 15 cm. The angle of the rotary stage with respect to the x-axis required taking 2D projection scans when the beam was placed close to the surface of the rotary stage. The apparent curvature of the resulting image was kept as low as possible. Any angular deviation of the z-axis leads to degraded resolution in the axial direction; however, the actual influence of a misaligned z-axis on the image quality was not further examined.
Alignment of the scan center was necessary to remove ring artifacts in the reconstruction. Because the edges of the steel rod can be detected in the projection scan by software, the scan center alignment was implemented in software: eight projection scans were taken at 45° increments, and the center position between the edges was determined. The average center position of the eight projections became the new scan center, and the difference of the center positions between alternating scan directions was used to compensate for mechanical slack in the x-axis actuator.
Imaging Performance and Image Quality
Mechanical constraints limit the sample size to a width of approximately 20 cm and a height of 15 cm. Depending on sample size and number of projections taken, the scan time for a single slice is between 7 and 15 min. The slow scan time is a consequence of the parallel-beam design imposed by the existing X-ray equipment, because it requires a mechanical scan motion for the acquisition of each projection. Acquiring multiple slices for a volumetric scan takes proportionally longer, and 3D scan times of several hours can be required with the present device. During acquisition, the tube dissipates approximately 50 W (alternating 70 and 140 kV for 4 ms during one AC cycle of 16.7 ms with a tube current of 1 mA) and therefore slowly heats up the oil in the X-ray tank. After several hours of uninterrupted operation, the tank tends to become hot to the touch (estimated 50°C), and forced-air cooling was later added by means of three 120 mm computer fans.
The scan of an aluminum stair-step wedge (Figure 5) allowed the system's signal-to-noise ratio and line-spread function to be measured. With artificially reduced pixel sizes of 0.2 mm (motor steps smaller than the beam diameter), the thinnest edge of the aluminum step wedge was imaged as an approximately linear gradient of 0.6 mm width, which is consistent with the 0.6-mm collimation aperture. With a Gaussian beam profile, a Gaussian line-spread function would be expected, but resolving the edge over 4 pixels makes it impossible to distinguish different models for the line-spread function. The important result is that the beam diameter is indeed the key factor that limits detail resolution. Moreover, during the acquisition phase of 4 ms, the motor advances by only 0.12 mm, and the resulting motion blur is lower than the beam diameter and therefore negligible.
Noise is introduced predominantly by the photomultiplier tube detector, and the noise component increases with the detector gain. If the signal-to-noise ratio is defined as the achievable contrast divided by the standard deviation of pixels in a homogeneous area, the device delivers projection scans with a 45-dB signal-to-noise ratio at the highest gain setting. Smaller or less dense objects can be imaged at lower gain settings, and the signal-to-noise ratio increases accordingly.
Image Value Calibration
The projection image of the stair-step wedge (Figure 5b) can be used to obtain the approximate value of the effective beam energy. Figure 6 shows averaged intensity profiles over the steps and the linear regression of base e absorption against aluminum thickness. The slopes in Figure 6b indicate apparent X-ray absorption coefficients for aluminum of 0.69 cm−1 for 140 kVp and 1.39 cm−1 for 70 kVp. By comparing these values to those published by the National Institute of Standards and Technology, effective apparent beam energies of 66 and 43 keV were found for the aluminum phantom.
Typically, CT scanners provide reconstructed images calibrated in Hounsfield units (HU), which define air as –1,000 HU and water as 0 HU. A linear relationship between the reconstructed absorption coefficient and Hounsfield units is assumed. For this reason, the scan of a simple water phantom provides the two required calibration values μwater and μair.7 In this special case, the detector gain setting influences the calibration in Hounsfield units, and the inclusion of a water phantom in each image allows the calibrated reconstruction even after acquisition. To provide an orientation of the values in this CT scanner, the reconstructed cross section of a 50-ml centrifuge tube filled with distilled water and imaged with the same parameters as the aluminum wedge had a mean value of 0.21 cm−1 for 140 kVp and 0.31 cm−1 for 70 kVp. A conversion of the aluminum absorption coefficients to Hounsfield units yielded approximately 2,300 HU at 140 kVp and 3,500 HU at 70 kVp. The dependency of the aluminum CT number on the beam energy is not unexpected8,9 and rather highlights the shortcomings of the Hounsfield unit system than those of the CT scanner.
Some example images that emerged from ongoing studies are shown in Figure 7. Cross-sectional images of an onion (Figure 7a and 7b) reveal the individual layers and the internal structure. These images are part of an investigation whether CT can be used to detect bacterial rot in onions. Volumetric images can serve as the starting point for patient-specific biomechanical studies. A volume rendering, created from a stack of 2D slices of a denture, can be seen in Figure 7c. The CT image reveals interior structures, such as defects or supporting metal inserts. From the image, a mesh can be generated (Figure 7d), which serves as the starting point for finite-element analysis of the load-bearing capacity. Figure 7e and 7f show the cross-sectional image of bone density phantoms with 50, 200, and 500 mg/ml bone mineral embedded in a resin (Computerized Imaging Reference Systems, Norfolk, VA). In the first case (Figure 7e), the phantoms were arranged around a water-filled glass vial and placed in a 50-ml centrifuge tube for imaging. In the second case (Figure 7f), the centrifuge tube itself was filled with water. Shown are images created from weighted sums of the 140-kVp and the 70-kVp image (70% and 30%, respectively). Correlation between Hounsfield units and mineral content (107 HU for 50mg/ml, 453 HU for 200mg/ml, and 1,094 HU for 500mg/ml) was extremely high, with an increase of about 2.2 HU per mg/ml mineral density and R2 > 0.999.
The underlying idea for this article was to use key components from surplus stock to build a very low-cost yet fully functional X-ray CT scanner. Clearly, the X-ray components (X-ray tube with high-voltage generator and X-ray detector) are the most expensive individual components in any custom CT project. Design of the final CT scanner therefore depends almost exclusively on the available components.
At one end of the spectrum, for example, a CT scanner could be built from selected modules, such as a modular X-ray generator, X-ray camera module, and three-axis motion stage, with the additional capability to either rotate the X-ray system around the sample or rotate the same inside the X-ray beam. On one hand, this solution offers the greatest design flexibility and allows to steer the design in the direction of desired imaging parameters, such as beam energy, resolution, or field of view. On the other hand, this solution arguably is the most expensive option, and total costs can reach those of lower-priced off-the-shelf scanners. Moreover, modular commercial components often come with their own software, and the integration of disjunct software control modules may pose a challenge.
Custom-built CT systems that follow this design principle have been reported in the literature. Two examples are the micro-CT system of Paulus et al.10 and the volumetric scanner by Ross et al.11 In both cases, the design focus was to achieve a combination of high speed with high resolution, and commercial components were used. A cost overview was not given.
Very slow acquisition may affect degradable samples, but even more importantly, it has a direct effect on the lifetime of the device because of the natural degradation of the X-ray tube due to material evaporation from the cathode.
At the other end of the spectrum, the high-value components are available at a low cost, for example, from surplus equipment. Only the sample motion unit and control for the entire system needs to be provided. The resulting costs likely are between one and two orders of magnitude lower. However, strict design limitations are imposed by the features of the available X-ray units. In our case, the tightly collimated beam allowed only a first-generation parallel-beam geometry with accordingly long acquisition times, which likely is the most noteworthy drawback of the present device.
Very slow acquisition may affect degradable samples, but even more importantly, it has a direct effect on the lifetime of the device because of the natural degradation of the X-ray tube due to material evaporation from the cathode. No information is available on the cathode temperature or filament current, but a crude estimate for an assumed tungsten cathode would place its operating temperature near 2,400 K to achieve a thermionic emission of 0.13 A/sq.cm.4 Such low-current tubes typically have long life spans of more than 10,000 hours. This estimate is consistent with the operation of the original DEXA device, for which the tube might have been energized for up to 2,000 hours per year. Unfortunately, no information about the history of the tube is available, which is a typical scenario for surplus parts. Therefore, estimating the remaining lifetime of the tube is impossible. On the other hand, net tube usage of only 260 h was logged over the past year, where a study on bacterial rot detection in onions was performed, and tube aging during lab usage appeared to be quite limited.
An additional constraint imposed by the existing equipment was the limitation of the spatial resolution to 0.5 mm by the given beam collimation. The travel distance of the translation motor per pixel could be decreased easily, which would lead to a higher apparent resolution. However, blur caused by the beam diameter would prevent details at the pixel size from being recognizable. Although further collimation with a pinhole aperture would be possible in principle, an unacceptable degradation of the signal-to-noise ratio can be expected.
To provide one hypothetical example for different design constraints, let us assume that the available components are a wide-field X-ray generator and a digital detector array. With the cone-shaped beam that the X-ray generator emits, much faster acquisition times are possible through fan-beam or cone-beam techniques. However, the tube's focal spot would be the primary factor that limits the resolution, and scattered radiation may cause a substantial noise component in the projection scans. Moreover, image reconstruction becomes more complex. For an idealized fan-beam geometry, geometrical correction terms can be introduced that map the fan-beam geometry to parallel-beam geometry and also that correct the frequency response of the reconstruction filter.12 Fan-beam reconstruction therefore is fundamentally similar to parallel-beam reconstruction. The complexity of the reconstruction algorithm grows dramatically with the cone-beam geometry. The algorithm that is probably most widely used is the iterative 3D generalization of the filtered backprojection by Feldkamp et al.,13 but a Fourier-based approach also exists.14
Radiation safety compliance is a fundamental aspect of the design, as radiation sources are regulated by federal and state laws far beyond radiation exposure levels.
After considering constraints imposed by the X-ray source and detector, however, design of the additional components becomes more flexible. In our design, we made use of the versatility of low-cost microcontrollers to control motion, X-ray generation, acquisition, and safety. Integrating some or all of these functions in software that runs on a personal computer is possible. Labview (National Instruments, Rockville, MD) is a popular software to realize control and automation systems. However, the Labview license and the necessary interface cards incur additional costs. Alternatively, programming the software in, for example, C or C++ is feasible. This approach allows easy integration with the reconstruction and display modules. The two main challenges in this case are precise timing and interfacing with the hardware. Interface cards can be purchased, for which software libraries are sometimes available. At this point, however, Labview and similar high-level programming systems allow faster software development. On the other hand, the use of microcontrollers provides unmatched design flexibility at extremely low costs. Microcontrollers provide integrated on-chip timers and analog-to-digital converters. Microcontrollers therefore are ideal for timing-critical tasks, such as step motor acceleration, or for data collection.
Radiation safety compliance is a fundamental aspect of the design, as radiation sources are regulated by federal and state laws far beyond radiation exposure levels. Radiation shielding usually is only one issue, with the choice being between a device that is operated in a shielded room (for easier access) or a cabinet-type device that is housed in its own shielded cabinet. Additional considerations include the safety interlocks, indicator lamps, and keyed access, as well as documentation, operator training, regular radiation surveys, and regular tests of the safety features that are necessary after the design phase is complete. Consultation with the institutional radiation safety division in the early stages of the design is recommended, and in our case, the radiation safety division at the University of Georgia was instrumental in identifying necessary steps to bring the device in compliance.
A number of options exist to further improve the device, but most of these options revolve around reconstruction. First and foremost, no consistent dual-energy reconstruction has been implemented for this device. Although dual-energy reconstruction is not a novel method,15–17 the different noise levels at 70 and 140 kVp pose a challenge18 for dual-energy subtraction images. On the other hand, implementation of dual-energy reconstruction is attractive because it allows the density of one material (e.g., water content) to be determined in a sample, as well as allows material contrast to be improved9 and beam-hardening artifacts to be suppressed.19 The CT device itself, rather than the more commonly used simulation methods, could be used to provide data for the evaluation of new methods for noise reduction20 or streak artifact suppression. 21 In a similar fashion, the influence of signal conditioning on the hardware side (e.g., use of a high-order antialiasing filter before the digital data collection) can be examined.
Furthermore, a conversion to a helical scanner22 is straight-forward with the current design: either the rotary stage is mechanically coupled with the z-translation stage or the z stage is programmed to advance incrementally after every projection scan. The second option particularly is attractive because it is a pure software solution and can be selectively enabled. The main advantage of helical scanning in this device likely would be accelerated 3D scanning, because a compromise can be found between axial speed and axial detail loss through interpolation.
In fact, notwithstanding the cost aspect, the open nature of a CT device designed and built from the ground up arguably is its most attractive aspect, because it enables implementation of, and experimentation with, different acquisition protocols and reconstruction techniques. A device such as the one described here is not only suitable for imaging studies at the quality level of commercial devices, but it also enables researchers to advance CT technology itself.
About the Author
Mark A. Haidekker, PhD, is a professor in the Driftmier Engineering Center, College of Engineering, University of Georgia, Athens, GA. E-mail: email@example.com