How does the ROI affect the thresholding? Micro-computed tomography can be applied for the visualization of the inner structure of a material or biological tissue in a non-destructive manner. Besides visualization, image analysis is an important topic in micro-ct, to obtain quantitative parameters from scanned datasets. SkyScan has developed analysis software CTAn to serve this purpose to a large extend. An important issue to consider in CTAn is the application of a region of interest (ROI). This application note will give an example to illustrate the importance of this ROI and the correct use of it. Intro CTAn performs quantitative measurements both of densitometry (voxel attenuation coefficient or calibrated density) and of morphometry. The morphometry image analysis is based on binary images. Binary images contain only black and white pixels, representing the non-selected or selected pixels respectively. The process of selecting pixels is called binarization and in CTAn this is performed by means of segmentation or thresholding. Simple global or adaptive methods allow for easy segmentation. Acquisition of quantitative data can be greatly influenced by defining the region of interest. CTAn offers a number of tools for drawing ROI s with a wide range of shapes. The importance of this ROI needs to be emphasized. Methodology Let s consider the following illustrative dataset, a sand sample. The dataset has been acquired with a SkyScan1172 with a 10MP Hamamatsu camera. The sample was scanned at 44kV with a 0.5 mm Al filter at 360 and reconstructed with NRecon. This 8-bit reconstructed dataset is loaded in CTAn and has 256 different greyvalues in the range [0,255]. 1
Figure 1. Illustrative sand sample in CTAn. The table below (Table 1) shows the different steps in 2 comparative methods. In the first method an ROI is drawn, a new dataset is saved and reloaded from this ROI, and 2 different thresholding values have been applied to select the air in between the sand particles. The second method follows the same steps, but there is an additional step in between. It not only loads the new dataset from the ROI, but also the ROI itself. Understanding the impact of this key step, and how this affects the results (in step 4 and 5) is crucial for analysis. Table 1. Comparison of 2 analysis methods Method 1 Method 2 Step 1: draw an ROI Step 1: draw an ROI 2
STEP 2: save new dataset from ROI, load (this) new dataset STEP 2: save new dataset from ROI, load (this) new dataset STEP 3: Load the ROI (again) STEP 3: Apply thresholding [1,25] STEP 4: Thresholding [1,25] STEP 4: thresholding [0,25] STEP 5: thresholding [0,25] 3
[0,25] [1,25] Result & discussion CTAn looks at an image as a square collection of pixels. When you are loading an image with a different shape, CTAn will automatically generate a square picture from this shape, by including a number of black pixels with grey value of 0. When reloading the dataset in step 2, the round image has now become square. When thresholding the range [1,25] these pixels are not included, but when the boundary of the thresholding range is changed to [0,25], these pixels are also binarized, showing that these pixels have a greyvalue of 0. In the second method, not only the dataset has been reloaded, but also the ROI. This is a way of defining the image boundary. The image is now round. The pixels in green are not part of the image, which is seen when applying the thresholding ranges [1,25] and [0,25]. The image is now defined within the boundaries of the ROI. This difference has a consequence in analysis. When analysing the percentage of air in this sand sample for example, one can be interested in the object percentage. In the table below, Table 2, this parameter is calculated for each of the four possible outcomes from Table 1. When applying a threshold of [1,25], the object percentage is 39.8%. When neglecting to reload the ROI, an artificially low value of 30.7% is obtained. Although the absolute amount of white pixels remains the same This value is lower because the total ROI area is larger. There are now less white pixels with respect to the ROI. For a thresholding value of [0,25], the object percentage would be 49.8%.. When the reloading of the ROI would have been neglected, this would lead to an artificially high value of 61.3 %. The pixels surrounding the round sand sample are now also included in the image, and considered as part of the object, which is obviously a false statement. Table 2. analysis of different ROI & thresholding Not reloading ROI ROI Area = 9.14 106 micron ² Reloading ROI ROI Area = 7.06 106 micron ² 30.7% 39.8 % 61.3% 49.8 % 4
Conclusion In this application note the importance of understanding the ROI has been pointed out. By means of an example dataset it has been shown that neglecting to reload the ROI may result in an inappropriate amount of pixels being falsely considered as part of the image, which can lead to incorrect analysis results. Obtained values can be artificially high or low, depending on the protocol that is used. In order to prevent this from happening, one needs to carefully address this issue and design an appropriate ROI. SkyScan, Belgium 5