Improving Segmentation of Breast Ultrasound Images: Semi Automatic Two Pointers Histogram Splitting Technique

Motivation of this step is to have estimating information about the lesion. Usually breast cancer lesions have deformed shapes and sizes  [22]. It is very difficult to categorize them into any common classes. To extract any feature from the lesion, we need to manually put a simple and uniform shape, such as a circle inside the lesion. An user puts two points, the center of the circle and a point on the circumference respectively, which has the radius (R) computed automatically from the coordinates, to generate this initial circle. Reason for having a circle instead of a square or rectangle or any other shape is that the trajectory of the second point is equidistant from the first, making it easier to estimate during selection. The circle should be within the lesion, covering as much as possible. Slightly overshoot-ed circles don’t show much varied result than the fully engulfed (within the lesion) one. This is due to the fact that in most of the cases, a small change in shape of the circle does not make big changes in its histogram outcome. Also, for most segmentation algorithms, we have to initialize many points. Here we manually initialize only the two points, the circumference, which is generated automatically, feeds the other points into the algorithms, which will follow later.

It is very likely that the pixel characteristics within the lesions are quite different than the pixels outside that area. Using this fact we use RP and LP to process the whole image to get a better filtered image. We call this filtered image Histogram Split and Stretched Image (HSSI). To get this HSSI we take all the pixel values that is right to RP on the histogram indices, xi¿xr, and multiply it with a RightSplittingFactor (RSF), which is greater than 1. The reverse is implemented on the pixel values of xi¡xl, that is left to LP, multiplying with a LeftSplittingFactor (LSF), which is less than 1. The splitting of these pixels have a meaningful effect on the real image, providing a new insight of the HSSI. Looking into the histogram in figure 1(g) and 1(h), we notice that the histogram is split and later on stretched depending on the RSF and LSF, hence the name of the image. It makes a blank region between the MCP and LCP, both in the histogram of within the circle and the whole image. Pixels at LCP change and this makes a significant relative difference between the pixels values within the ROI and outside the ROI (Figure 1(e) ). The edge image of this HSSI at figure 1(f) has better edge image. In most cases it is clear of many edges within the lesion area. The difference between the edge image of the original BUS image in figure 1(b) and the edge image of the HSSI (EHSSI) in Figure 1(f) is clearly visible.

To determine how classical segmentation algorithms perform on with respect to original image, we used active contour snake model  [14], Metamorph segmentation  [16], Regular Level Set and Distance Regularized level set evolution segmentation model  [15]. All of these models take in some initialized points, to estimate ROI. In this case, we use the unmodified version of these algorithms on both the original image and HSSI with the same initialized points. After the segmentation we obtain the border and derive the binary image from it, white being the object, black the background.

This is a work in progress, which shows statistical improvement in the process, which may be used in the future to develop segmentation algorithms just from the edge image. Looking into the edge image of the HSSI (EHSSI) we can see the improvement when compared with the original edge image. Split histogram made the lesion area more saturated, leaving less edge pixels, and more noise outside the ROI, adding to previous edge pixels. We refer “edge pixels” to those pixels which appear in as pixels in edge image. These change in the edge pixels help the segmentation algorithms perform better, as most of the algorithms, take edge map of an image as an initial parameter to find the ROI. One probable target of work in the future can be to only use this edge map to segment out the ROI of the original image. For this, we perform two operations on the EHSSI –

1. Edge wash-up: It is obvious that the center of the circle must be in the lesion. To remove unwanted edge pixels that appear very close to the center we wash up all the edge pixels in a small radius (Edge wash-up radius, EWR) from the center of the circle. The size of EWR, depends on the radius (R) of the initialized circle, proportionally. The obtained image is called “Washed-up” image. We perform our pixel irregularities test on the result section to find out more about image.

2. Imclose operation: To fill up the holes, we perform flood-fill operation in the background (imfill in Matlab image processing toolbox) followed by dilution and erosion and lastly flood-fill again. This process tends to make the outside of the ROI look like a solid background, we call this image Enhanced EHSSI or EEHSSI. We will follow up the results of the precision of EHSSI in the result section.

All the ground truth of the BUS images were manually delineated by expert clinicians. Comparing the segmented regions by the algorithms with the ground truth image we can see improvement in the results. As a comparison parameter, we used Sorensen Dice Similarity Co-efficient (DSC). Many previous works such as , [10]  [1],  [Pons2013] etc. have used the following DSC formula to determine how precise the segmentation is with respect to the manually delineated one. If the segmented binary image is X (fig. 3(c)) and the ground truth is Y fig. (3(d)) , then, DSC parameter

Segmentation algorithms of the study were collected from various source. To compare the results with active contour, we used Matlab inbuilt function activecontour [Matlab copyright 2012-2017] for the active contour segmentation. For MetaMorph segmentation, we used the codes provided by the authors themselves. Two level set methods, Distance Regularized Level Set Evolution (DRLSE) and Spacial Fuzzy Cluster and Level Set Segmentation (SFCLS) were collected from MathWorks File Exchange, uploaded by the author himself for DRLSE at  [15] and SFCLS by “ABing” from National University of Singapore in  [23].

Comparing with different segmentation algorithms, we get the following results:

A thing to note is that the results of segmentation of the algorithms may seem to lag the original claimed quality. The reason is the initialization. Most algorithms require more precise initialization, where as our method only requires a not-very-precise simple circle to initialize.

Pixel irregularities test is a study performed on the lesion area of the edge maps obtained from the original image and the HSSI (EHSSI). In the lesion area (fig. 4(c)), ideally there should be no edge pixels, because in ideal case, that region should be homogeneous. But practical ultrasound diagnosis ends up with low-resolution, non-homogeneous regions, yielding edge pixels in the lesion area. HSSI provide us with more homogeneous region in the ROI. To compare the two edge images fig. 4(c) and (d), we perform test to find the percentage of wrong pixels appearing in the lesion area of the edge map. This percentage is the ratio of wrong pixels appearing in the ROI and the area (by pixels) of the ROI by hundred, denoted by Pixel Irregularities Ratio (PIR). Area of the ROI was obtained from the ground truth.

It was also performed on the washed-up image (fig. 4(e)) with better result outcomes. As we know, the less the wrong pixels in the ROI, the lower the PIR, the better the result, showed in Table 2. We also calculated the number of images where our method yields better edge map inside the ROI.

At worst cases, our method can keep the system as it is. If the lesion area is very random in nature, then the two pointers are not very helpful to distinguish the MCP and LCP. So, the algorithm is unable to generate better HSSI image than the original one. Figure 4 shows the comparison of images where our method fails to improve the results.