Porosity determination in additively manufactured Ti parts using X-ray tomography

Ti6Al4V is a suitable titanium alloy for all kinds of medical implants and prostheses because of its high durability and biocompatibility. Furthermore, components of high complexity can be produced via additive manufacturing which allows for more flexibility and easy prototyping of patient specific implants. However, this flexibility implies the risk of internal defects resulting from the manufacturing process. The nondestructive investigation of critical components is therefore crucial to avoid premature failure. X-ray micro-computed tomography (XCT) is a method that can resolve internal structures three dimensionally in a non-destructive way. Nevertheless, the probability to detect defects is limited by the achievable resolution and image quality of a scan. In this contribution, we performed a systematic study to determine the pore size distribution in additively manufactured Ti6Al4V parts using XCT. We focused on the influence of scanning parameters such as voxel size, tube voltage and current on the image quality that determines the outcome of the porosity analysis. Image quality was assessed via contrast to noise ratio (CNR) and slanted-edge modulation transfer function (MTF) according to ISO 12233. Furthermore, we optimized the beam hardening correction for the scans and investigated influences of different image denoising algorithms. Results showed that tube voltage and current greatly influence the CNR of the data set while the MTF is, within limits, almost constant as long as the electron beam focus is optimized. With higher physical resolution, smaller defects can be detected, which leads to porosity values of 0.36, 1.35 and 2.54% at 10, 5 and 2.5 µm resolution respectively. Image post-processing can further influence porosity outcome because of the segmentation of noise induced particles. Different image denoising algorithms therefore can heavily reduce porosity values depending on spatial resolution


Introduction
Titanium and its alloys are among the materials most frequently used in biomedical applications due to their high strength and biocompatibility at moderate weight.Nevertheless, because of high manufacturing costs and rather poor workability the usability of titanium for complex structures is limited.With the development of additive manufacturing (AM) methods such as selective laser melting (SLM) this drawback can be compensated with the ability to produce complex geometries layer by layer from scratch.However, AM parts often suffer from porosity introduced during manufacturing that affects mechanical properties.Consequently, it is advisable to perform non-destructive testing (NDT) by X-ray micro-computed tomography (XCT) to identify defects in safety critical components.In comparison to specimen preparation and microscopic evaluation, 3D image data of internal structures is acquired and analyzed non-destructively using XCT.However, a trade-off between spatial resolution and sample dimensions is often unavoidable, which can lead to an inaccurate determination of defects in the material.In order to perform an optimal XCT scan also for experienced measurement engineers it is often necessary to perform several test scans, since there are numerous factors that can influence the qualitative outcome of a scan, and slight changes can cause significant differences.In this work, we therefore empirically investigate the influence of scanning parameters including voxel size, tube voltage, and current on the outcome of XCT scans performed on additively manufactured Ti6Al4V parts.The effect of spatial resolution on the detectability of pores and total porosity can thereby be assessed similarly to a previously performed study for carbon fiber reinforced polymers [1].

Materials and methods
We systematically investigated a cubical specimen of Ti6Al4V, additively manufactured by SLM at the Center for Smart Engineering (CSM) at the University of Applied Sciences Upper Austria (Campus Wels).In the following chapters, the preparation of the sample as well as scanning parameters and evaluation will be described in detail.

Sample preparation
The 10x10x10 mm specimen was produced on a Concept Laser M2 Cusing device (Concept Laser GmbH, Lichtenfels Germany) operated at a laser-power of 100 watt and 800 mm/s speed, moving in a chessboard pattern at a track width of 0.16 mm and a layer thickness of 30 µm.The grain size of the Ti6Al4V powder used is 20 to 63 μm; after production, no additional heat or surface treatment was performed.For the XCT investigation at smaller voxel sizes a section of 3x3x10 mm was cut out of the center of the cube by wire eroding as shown in Figure 1.

Data acquisition
Specimens were scanned using a Nanotom 180NF (GE phoenix | X-ray) system at isometric voxel sizes ranging from 2.5 -10 µm.To investigate the influence of scan parameters on image quality, scans were performed repeatedly at varying tube voltage and current while target power was kept constant for comparability of results.Target power was chosen in order to maintain a focal spot blur below the detector pixel pitch according to Weiß et al. [2].Since the theoretically applicable power at the 10 µm scan exceeded the XCT systems capabilities, two scan series at power levels within the typical operation range of the system were performed.Additionally, two different pre-filter plates were used for all series.Voltage levels (U) within a scan series were chosen to vary roughly 10 kV between each scan while the current (I) was adjusted to maintain a consistent target power (Pt).Voltage range was established to cover the full range between an oversaturation of the detector at high voltage and a minimum required transmission at low voltage levels.With these settings, a total number of 28 scans were performed as continuous scan at 700 ms exposure time and 1600 projections, reducing the scan time to 19 minutes per scan.Voxel size has been adjusted using a calibrated ruby ball bar (GE Sensing & Inspection Technologies GmbH-phoenix/x-ray).The cube was tilted by 45° during scanning to minimize Feldkamp artifacts at specimen edges.Cutout sections were scanned upright to maximize resolution.The complete scan settings are visible in Table 1.

Data analysis
To evaluate optimal scan settings, image quality was assessed via contrast to noise ratio (CNR) and slanted-edge modulation transfer function (MTF) according to ISO 12233:2000.CNR was evaluated in VGStudio MAX 3.1 (Volume Graphics) via Eq. 1.
The material region was defined over the complete specimen volume excluding pores.Air region was defined as the complete surrounding, excluding any parts of specimen fixation such as glass rods, glue or markers.MTF was calculated on a surface edge in a central region of the specimen averaged over 100 slices to reduce effects of noise and surface irregularities using a Matlab script by P. Burns [3].For comparability, within each series the exact same region was evaluated.
̅̅̅̅ mat … mean grey value in material region ̅̅̅̅ air … mean grey value in air region mat … standard deviation in material region Subsequently, data sets with the best image quality were corrected for beam hardening by applying different strengths of beam hardening correction (BHC) and comparing the resulting entropy in a single slice image as explained in [4].With optimized BHC level, entropy should be a minimum.Nevertheless, at 2.5 µm scans this assumption was no longer valid since the image noise also increases with the strength of the BHC.As a result, lowest BHC levels always yield minimal entropy.Therefore, an alternative simple approach for the optimization of the BHC level was implemented, which minimizes variances in a grey value line profile through the material.This avoids cupping effects of beam hardening and eventual overcorrection (see Figure 2).As a comparison, Figure 3 shows the uncorrected 2.5 µm scan on the left and the BHC corrected scan on the right.In this case, entropy is lowest at the uncorrected data set on the left, while the optimization approach implemented by use of the grey value line profile yields the best results at higher BHC strengths.It is clearly visible that although noise and therefore entropy increases with BHC, beam hardening artefacts as well as contrast are improved in the corrected data set on the right.

Results
Since the XCT scan parameters tube voltage, current, and pre-filtering can significantly influence the qualitative outcome of a scan, the first part of this study aimed at the optimization of these parameters.In the second part, effects of different postprocessing steps on the resulting porosity values are investigated.Optimal scan parameters were assessed by an analysis of CNR and MTF of each scan.While CNR can be represented by a single value, the MTF is a characteristic curve representing the remaining contrast depending on line pair frequency.Figure 4 shows exemplary MTF curves of the scan series with 2.5 µm voxel size and 0.1 mm copper pre-filter.
Figure 4: Slanted edge MTF of the 2.5 µm scan series with 0.1 mm copper pre-filter applied.
For simplicity of visualization and comparability, the result of each slanted-edge MTF was reduced to a single value calculated as the average of the MTF curve from 100% contrast to the estimate point where contrast reaches zero (0.2 line pairs/pixel at 2.5 and 5 µm voxel size, and 0.4 at 10 µm voxel size).Both CNR and MTF were normalized to the maximum occurring value at each voxel size and are given in Figure 5.  Since image quality directly influences porosity results, values increased by up to 66.67, 6.38 and 5% in median filtered 10, 5, and 2.5 µm scans respectively, depending on the scan parameters.In addition to the CNR and MTF, the minimum transmission was recorded for all scans performed.The relationship between CNR and minimum transmission similar to a study performed by Reiter et al. [5] are shown in Figure 6.The optimal scan of each series, i.e.where CNR and MTF are highest, was used for the following investigation of porosity.
Resulting differences in porosity evaluated via the same segmentation and post-processing pipeline can be seen in Figure 7. Furthermore, the influence of different post-processing filters on porosity results are also indicated.In Figure 8, a zoomed section of a pore with average size (72 µm diameter) is depicted to show the effect of voxel size on pore detectability.
Figure 7: Evaluated total porosity at 10, 5, and 2.5 µm voxel size, without post-processing (blue) and with median (orange) and non-local means filter (grey) applied.Note that the scan with 10 µm voxel size is performed for the complete cube specimen.Note that the scan with 10 µm voxel size is performed for the complete cube specimen (see Figure 1).
A more detailed depiction of the MTF of the optimized, NLM filtered 10, 5, and 2.5 µm scans is shown in Figure 9.For a comparison between different voxel sizes, the MTF was converted from line pairs per pixel to the edge distance between a line pair via the given voxel size.

Discussion
We showed that scanning parameters including voxel size, applied tube voltage, and current as well as different post-processing steps greatly influence the determined porosity value in additively manufactured Ti6Al4V specimens.This highlights the need for a systematic approach to optimize image quality since mechanical properties of AM parts are largely influenced by porosity.
Results of a previous study [6] revealed the importance of porosity for mechanical stiffness (e.g.Young's modulus) on AM titanium parts.Only accurate volume data can provide sufficient detail for finite element simulation models that are necessary for the improvement of additively manufactured Ti6Al4V parts in primary structures.The investigation of varying scan parameters was conducted in a way that the complete range of possible voltages was covered from a minimum required transmission until grey values close to an overexposure of the detector.From Figure 5 it is visible that the CNR tends to peak towards the center or center-right of this range while the extreme high or low voltages result in the lowest CNR.It needs to be noticed that at the 0.4 mm copper pre-filtered 10 µm scan this observation does not apply and the highest voltage results in the best CNR.The reason for this is that at the 10 µm scans the whole Ti6Al4V cube had to be penetrated by the X-rays, which resulted in generally low minimum transmission values of 6 to 10%.Therefore, more power would have been necessary for the penetration of the specimen, but at the same time, the detector was already close to overexposure.The MTF as a second parameter varies in a range of roughly 5% with exception of the 10 µm 140 kV & 2.5 µm 90 kV scans, which make a drop of roughly 10% in MTF.The reason for this might be the extreme settings of the scan, once at the high end, close to an overexposure of the detector, and once at the low end too close to a complete absorption of the X-rays.Especially in the 5 µm and 2.5 µm scans, the relatively low fluctuation in MTF might be in the range of error caused by inaccurate focus optimization.Consequently, one scan was performed at 2.5 µm resolution without focal spot optimization, to demonstrate the importance of accurate focusing.The averaged MTF value for this scan was reduced by 20.4% in comparison to the focal spot optimized scan.We found that image quality, influenced by the scan parameters, affected the total porosity outcome by up to 66.67, 6.38 and 5% in median filtered 10, 5, and 2.5 µm scans respectively.Therefore, with suboptimal scanning parameters, total porosity values are increasing because of a decrease in CNR.Thus, additional noise induced "particles" are segmented as pores.However, the influence of scanning parameters on porosity decreases with increasing resolution.We were also able to confirm the findings from Reiter et al. [5], that minimum transmission solely is no sufficient criterion for optimal scan performance.Only in the 10 µm scans of the whole cube specimen, a tendency to a better CNR with higher minimum transmission was notable, which can be explained by the generally low minimum transmission at these scans.
The investigation of Ti6Al4V specimens scanned at voxel sizes of 10, 5 and 2.5 µm showed an increase in total porosity at smaller voxel sizes.Generally, with higher physical resolution smaller defects can be detected, leading to higher porosity values.The application of image denoising filters leads to a reduction in porosity values, which is caused by noise induced pores.
Particularly at the 10 µm scan, porosity is more than six times higher without filtering.A voxel size of 10 µm therefore is obviously too big for the presented inspection since the average pore diameter of 72 µm evaluated by the 2.5 µm scan is too small for the resolution capabilities of the 10 µm scan.The MTF of the 10 µm scan shows that the contrast is already reduced to 4.85% at 72 µm edge distance (Figure 9).This also becomes more obvious in Figure 8, which shows a pore with average diameter completely disappearing in image noise at 10 µm resolution.At higher resolution, the MTF shows that at 72 µm edge distance 66.5 or 73.3% contrast is remaining at the 5 and 2.5 µm scan respectively.Nevertheless, at smaller voxel sizes, total porosity is still reduced after filtering since information about small details and pores are lost due to blurring effects.However, also the segmentation of noise induced pores is decreased, which makes filtered data preferable, especially since the edge preserving non-local means filter further reduces the loss of fine details.Summarizing, we have presented that preliminary studies for the optimization of scanning parameters are highly recommended and minimum transmission solely is no sufficient criterion for optimal scan performance.Furthermore, attention should be given to post-processing such as noise removal as it will greatly influence final results of porosity investigations.

Figure 1 :
Figure 1: AM Ti6Al4V cube with cutout region indicated red (left) and wire eroded cutout section for high detail scans (right).

Figure 2 :
Figure 2: Grey value line profile through the specimen.The red area between smoothed line profile and maximum value (max.) is minimized.

Figure 3 :
Figure 3: Comparison of an uncorrected (left) and BHC corrected scan (right) at 2.5 µm voxel size.To investigate the effect of noise filtering on porosity results, two different noise reduction algorithms were applied to the BHC data sets.A 3x3x3 median filter in VGStudio MAX 3.1 and a non-local means (NLM) filter (search window: 30, local neighborhood: 4, similarity value: 0.5) in Avizo (Thermo Fisher Scientific Inc.).Subsequently, porosity values were calculated in VGStudio MAX 3.1 by histogram-based global thresholding with a grey value threshold set to 50% between the air and material grey value peaks (ISO50).The minimum pore size was set to 27 voxels in order to exclude noise induced pores from the evaluation.

Figure 6 :
Figure 6: CNR as a function of minimum transmission.Scan voxel sizes are displayed in different colors and copper pre-filter plate thicknesses with different indicators.

Figure 8 :
Figure 8: Pore with average diameter of 72 µm at 10, 5, and 2.5 µm voxel size from left to right.Note that the scan with 10 µmvoxel size is performed for the complete cube specimen (see Figure1).