• Keine Ergebnisse gefunden

Majority classification

7.1 Preprocessing Results

Bias Field Correction

Bias fields due to magnetic field inhomogeneities were observed in several of the avail-able MRIs but were corrected to a great extent using the popular N4ITK [210, 211]

algorithm (Fig. 7.1). The first column in the figure represents the original image slice, the second the corrected image, the third the color-coded bias field as overlay, and the last column the difference of the corrected and original image. The color-coded bias field highlights the under- (blue/purple) and overexposed (yellow/red) regions in the images which can appear in varying forms and locations in the image.

Case 1 shows underexposure on the bones, the structure of interest for this work, in contrast to higher intensities of the fat tissue. BFC effectively restores the balance (second image in top row). Case 2 depicts a large underexposed region at the top of the image which affects the shaft of the Femur (see first and third image). N4ITK corrects the intensities in the region to a great extent, making the femoral shaft and Patella more visible. Case 3 is a good example for intensity inhomogeneity within bone structures which is vastly restored with the algorithm. Case 4 shows large intensity differences between Tibia and Femur (red area in bias field). The difference is corrected with BFC, which enhances the grey values of the Femur and diminished the ones of the Tibia. Case 5 shows extensive underexposure of the bones, especially in the shaft region of the Tibia. Here, the BFC enhances tissue contrast in the image and partially restores the visibility of the tibial shaft.

In addition to intensity inhomogeneities, MRIs of this work were also affected by arte-facts and BFC was only able topartiallycorrect these images (Fig. 7.2). The N4ITK algorithm assumes a noise-free scenario and can thus not remove MR artefacts. Nev-ertheless, underexposed regions in the images are enhanced and the differentiation between different tissues improved. Case 1 shows a noisy and underexposed image.

The subject had a large circumference of the subject’s leg and therefore, no knee

Case1

Original Corrected Bias Field Difference

Case2Case3Case4Case5

Figure 7.1: MRIs affected by intensity inhomogeneities were corrected using N4ITK [210, 211] algorithm. Original, corrected, color-coded esti-mated bias field as overlay, and difference of corrected and origi-nal image are shown from left to right for several cases. The color-coded bias field highlights under- (blue/purple) and overexposed (yel-low/red) image regions representative of the inhomogeneities.

7.1 Preprocessing Results

Case1

Original Corrected Bias Field Difference

Case2Case3Case4

Figure 7.2:MRIs affected by intensity inhomogeneitiesand MR artefacts. The N4ITK [210, 211] algorithm assumes a noise-free scenario and can hence not remove image noise (Case 1), wrap-around (Case 2 and 3) or motion artefacts (Case 4). Nevertheless, the intensities are par-tially corrected to improve the homogeneity within a tissue.

coil could be used which led to the degraded quality of the MRI. BFC does not remove the noise in this case but is able to improve the contrast in the image. Case 2 depictswrap-aroundartefacts, i.e. overlapping structures. In Case 3 there is a tiny circular region with extremely high intensities which causes the rest of the image to be underexposed. Additionally, a wrap-around artefact can be seen on the left side of the image overlapping with the Femur and other tissue. N4ITK cannot remove the wrap-around artefacts but is able to correct the induced inhomogeneities in the image to a great extent. Case 4 shows streaking artefacts due to motion which even affect the growth plate of the tibia. Again, BFC is not able to reverse this effect but improves tissue contrast due to field inhomogeneities.

The impact of different image sizes and spacings on the execution time of N4ITK was recorded (Table 7.1). Reducing an image from it’s original size of 512×512×24 voxels to 448×448×24 voxels (approximately 25%) translated directly to the reduction of the algorithm duration by 25%. A decrease in image spacing, e.g. from 0.56×0.56 mm2to 0.45×0.45 mm2, generally led to an increase of the execution time. The algorithm is dependent on the central processing unit (CPU) since it operates on multiple threads. This has to be taken into account given the reported values. The hardware details of the workstation available for this work can be found in Appendix A.

Table 7.1:Impact of image size and in-plane spacing on the execution times of the N4ITK [211] algorithm for bias field correction of 3D MR images Orientation State Image size Image spacing Execution time

(mm3) (min:sec) coronal original 512×512×24 0.39×0.39×3.9 06:01 coronal resampled 448×448×24 0.45×0.45×3.9 04:30 sagittal original 864×864×50 0.17×0.17×2.2 38:15 sagittal resampled 448×448×50 0.34×0.34×2.2 9:47 coronal resampled 448×448×24 0.45×0.45×3.3 04:00 coronal resampled 448×448×24 0.47×0.47×3.3 03:03 coronal resampled 448×448×24 0.49×0.49×3.3 02:26 coronal resampled 448×448×24 0.56×0.56×3.3 02:08

7.1 Preprocessing Results

Automated Cropping

Automated cropping based on patch matching was used to extract a standardized VOI in each 3D MRI to compensate for differences in leg position of the subjects during the MR examination and for differences in FOV (section 4.3). Next, a few examples showing high variance in the position of the knee joint in coronal and in sagittal images (Fig. 7.3, Fig. 7.4). The characteristic region, orpatch, is depicted in the first column, the original image in the second column, the resulting correla-tion map between the patch and the image in the third column, an the generated standardized VOI in the fourth column.

For coronal knee MRIs, the selected patch shows the interconyloid eminence (Fig. 7.3) and for sagittal MRIs, it is a different patch showing the posterior cruciate ligament (Fig. 7.4). In all the cases represented in the figures, the algorithm success-fully detects the best fit of the corresponding patch in the image and is color-coded as red in the correlation map. Finally, the examples show, how automated cropping enables the extraction of a standardized VOI, irrespective of the position of the knee joint and the selected FOV during the MRI examination.

The execution time of automated cropping, i.e. patch matching followed by the extraction of the standardized VOIs, was tracked for images with different spacings, number of slices, and orientation (Table 7.2). The duration was similar for coronal and sagittal MRIs, although the latter had higher spacing. Analysing the orienta-tions separately, the execution time was dependent on the spacing of the image. The lower the spacing, the higher the automated cropping duration.

Table 7.2:Execution times of the automated cropping step for two images with different spacing and size

Orientation State Image size Image spacing Execution time

(mm3) (sec)

Case1

Intercondyloid eminence

Patch Original Correlation Map Standardized VOI

Case2

Intercondyloid eminence

Case3

Intercondyloid eminence

Case4

Intercondyloid eminence

Figure 7.3: Extraction of standardized VOIs in coronal knee MRIs using a method based on patch matching. A patch showing the intercondy-loid eminence is slid across the original image and the best position matches the red area in the correlation map. This point is the cen-ter of the resulting VOI. The automated cropping generates similar VOIs regardless of the position of the knee joint (Case 1 and 2) and independent of the FOV (Case 3 and 4).

7.1 Preprocessing Results

Case1

Posterior Cruciate Ligament

Patch Original Correlation Map Standardized VOI

Case2

Posterior Cruciate Ligament

Case3

Posterior Cruciate Ligament

Case4

Posterior Cruciate Ligament

Figure 7.4:Extraction of standardized VOIs in sagittal knee MRIs using a method based on patch matching. A patch showing the posterior cruciate ligament is slid across the original image and the optimal po-sition is color-coded as red in the correlation map. This point is the center of the resulting VOI. The automated cropping generates sim-ilar VOIs regardless of the FOV (Case 1 and 2) and independent of the position of the knee joint (Case 3 and 4).