6. PROPUESTA DE CAMPAMENTO 43
6.4. Descripción de las actividades
With the preceding datasets and the visual comparison between X–ray mam- mogram and projection of the speed of sound image, the cropping of the
Figure 4.9: This figure shows the X–ray mammogram and the projection of the speed of sound image of dataset 4. Both images render large amounts of glandular tissue. On the left side of the breast is a tumor of the size 19x15x22.
X–ray mammogram can be examined. As described in chapter 3.3.2, the po- sition to crop the X–ray mammogram is determined by scaling the projection of the speed of sound image to match the contours of the X–ray mammogram and assessing the similarity between X–ray mammogram and scaled projec- tion. This similarity is measured by NMI (see chapter 2.2.1). To achieve the best NMI, the projection of the speed of sound image needs to render tis- sue structures similar to the structures rendered in the corresponding X–ray mammogram. The adjustments of the segmentation described in chapter 4.2 supports a proper rendering of these tissue structures.
The first test is carried out with dataset 3 as the tumors are visible in both, X–ray mammogram and projection of speed of sound image.
The X–ray mammogram of dataset 3 has a length of 668 pixels (235 mm). According to chapter 3.3.2, the first position to crop the X–ray mammogram is at 70% of its length. 70% of the length equates to 467 pixels (164 mm). After comparison with the scaled projection of the speed of sound image, the length of 467 pixels delivers the best NMI. This length is the estimated
position where X–ray mammogram and speed of sound image of dataset 3 render same amounts of tissue.
Figure 4.10: The original X–ray mammogram of dataset 3 (left) is cropped (middle) to show the same amount of tissue as the corresponding speed of sound image. The position is estimated by the comparison with the scaled projection of the speed of sound image (right). In this projection, the second tumor can still be recognized (red rectangle).
Figure 4.10 shows the results of the automatic cropping. The cropped X–ray mammogram shows only one of the two tumors in its entirety. However, in the scaled projection of the corresponding speed of sound image the second tumor is still rendered.
The ultrasound of the Karmanos Cancer Institute records less tissue than X–ray mammography, hence, the cropped X–ray mammogram still has to show same tissue structures. In this instance, only one tumor is completely rendered in the cropped X–ray mammogram. Therefore, the automatic crop- ping delivers an inaccurate result.
To ascertain if the automatic cropping of the X–ray mammogram delivers in- correct results, this process is carried out with the remaining datasets. The diagnosis of the automatic cropping shows that the comparison between X– ray mammograms and the scaled projections of the speed of sound images always deliver the best NMI for the smallest amount of tissue rendered in the X–ray mammogram. This result is independent of the starting position to crop the image.
As described in chapter 4.3.3, the X–ray mammogram of dataset 3 shows a large breast with a large amount of breast tissue. When cropping at 70% of the X–ray mammogram’s length, the cropped X–ray mammogram shows a very similar amount of tissue as the corresponding speed of sound image. In the case of smaller breasts as in dataset 1, 2 and 4, the cropping at 70% of the X–ray mammogram’s length results in major differences in the amount of breast tissue between X–ray mammogram and speed sound image.
This inaccurate cropping also impacts the determination of the thickness of the breast during X–ray mammography, as the calculation is based on the cropped X–ray mammogram. If too much tissue is cropped in the image, the dimensions of the breast shown in the X–ray mammogram considerably decrease in size. This results in a higher estimated thickness.
As all further calculations are dependent upon the accurate cropping of the breast, this automatic process has to be modified.
A first analysis shows that the segmentation provided by the Karmanos Can- cer Institute underestimates the contour of the breast in the X–ray mammo- gram. In figure 4.11 a small white point is visible in the original X–ray mammogram. This point, a so–called “BB” belongs to a lead ball which is used to mark the nipple of the patient. The tissue between this ball and the visible glandular tissue is not segmented accurately. In the corresponding segmentation this part of the breast is missing.
The first attempt to improve the automatic cropping of the X–ray image is hence a new segmentation of the original datasets. Based on the assumption that both, the X–ray mammograms as well as the speed of sound images were segmented inaccurately, a new segmentation is carried out for images of both modalities. The new segmentation for the X–ray mammograms results in X–ray images which show the entire tissue from chest wall to nipple (see figure 4.12).
In the segmentation of the speed of sound images provided by the Karmanos Cancer Institute, the contours of the segmented breast were also underesti- mated as the area around the nipple was not shown. This is due to the fact
Figure 4.11: This figure shows the original X–ray mammogram before seg- mentation (left). The nipple of the breast is marked with a small lead ball (magnifier). In the segmented X–ray mammogram on the right, this lead ball is no longer visible.
Figure 4.12: The left image shows the segmented X–ray mammogram pro- vided by the Karmanos Cancer Institute. On the right, the new segmentation of the X–ray mammogram clearly renders a larger amount of breast tissue after the new segmentation.
that the slices are omitted which render the nipple since the artifacts are very dominant in this area. The new segmentation approach also segments the area around the nipple.
4.4.1
New Segmentation of Speed of Sound Images
The new segmentation of the non–segmented speed of sound images, roughly segments the background, glandular tissue and fatty tissue. After this seg- mentation, the contour of the breast is determined for each slice and the transition to the adjacent slices is smoothed. Afterwards, the entire three– dimensional contour is smoothed and the segmented speed of sound image shows the complete recorded breast with a tendency to overestimate the breast volume.
This process leads to a higher number of slices used for the segmented three– dimensional speed of sound images as the slices containing the nipple are no longer omitted.
The new segmented images were tested with the cropping method. Unfor- tunately, the automatic cropping still delivers inaccurate results. Further research of the segmentation has to be carried out. As this is beyond the scope of this diploma thesis, the position to crop the image is estimated man- ually. These lengths are 8.8 cm for dataset 1, 11.8 cm for dataset 2, 14.3 cm for dataset 3 and 12.1 cm of the breast for dataset 4.