General Overview:
Osteoarthritis is the most common joint disorder typically presenting its first symptoms in middle-age adults and can severely impair mobility. We are working to automatically segment the quadriceps muscles and the meniscus in the knee in order to aid clinicians in their study of the development, causes and effects of osteoarthritis.

Automated Segmentation of Quadriceps Muscles in MR Images
While the shape and size of the quadriceps muscles between different subjects varies widely, there is enough similarity to allow for the successful segmentation of the muscles using templates. The templates are thigh scans for different representative subjects and differ in the ratio of muscle to fat. The images below are two examples of template images used.

Figure 1 : Templates used for segmentation of thigh images

The image to be segmented is compared to the available templates by using the Kullback-Leibler divergence measure, DKL, which gives a value for the similarity between two distributions, in this case the normalized histograms of the thighs in the images (Figure 2).

Figure 2 : Comparison of the base image to the two templates using the Kullback-Leibler divergence between the normalized base image histogram and the normalized template histograms. It can be seen that template 1 has the lower divergence, and so is chosen as the image more similar to the base image, and hence used for the segmentation.

After the template is chosen, an affine transformation is calculated using point correspondences automatically extracted from the template and base images (Figure 3).

Figure 3 : Five fiducial points ( colored ‘x’ ) showing point correspondences between two different MRI scans (a) and (b).

The template has associated manual segmentations of each of the quadriceps muscles. These segmentations are registered to the base image using the affine transformation (Figure 4).

Figure 4 : Base images with template contours in red, showing (a) initial alignment and (b) registered alignment.

The base image is then processed to remove inhomogeneities, smooth constant regions (such as muscle), and enhance edges. Field inhomogeneities are removed using a homomorphic filter (Figures 5, 6).

Figure 5 : Base image (a) before and (b) after homomorphic filtering. The brightness of image (b) and the artifacts are enhanced by the scale used for visualization.

Figure 6 : (a) Fat ROI used for histogram calculation. (b) Prior to homomorphic filtering, entropy = 7.3593 (c) After homomorphic filtering, entropy = 4.1296.

Large regions of similar intensity are smoothed and edges between regions are sharpened simultaneously through the use of an anisotropic diffusion filter (Figure 7).

Figure 7 : (a)(c) Original image with gradient map, and (b)(d) anisotropically diffused image. Notice the smoother muscle regions and sharper edges in (d).

Finally, the registered template contours are evolved to segment the muscles in the base image using a level set approach (Figure 8).

Figure 8 : Final segmentations of the four quadriceps muscles. The overlap ratios between the automated segmentations and manual segmentations, using the Zijdenbos similarity index (with a range from 0-1, 1 being perfect agreement and 0.7 being “excellent” agreement), were:
rectus femoris – 0.87, vastus lateralis – 0.93, vastus intermedius – 0.86, and vastus medialis – 0.74

Semi-Automated Meniscus Segmentation
We have developed a semi-automated approach to segmenting the meniscus in order to reduce the inter and intra-reader variability associated with manual segmentations. Such an approach allows for a more accurate longitudinal assessment of meniscus changes with OA progression. The images on the next page demonstrate the improved accuracy of our program in comparison to manual segmentations. The meniscus is the darker region located in the center area of each image. Our algorithm for semi-automated segmentation (in yellow) provides a tighter fit around the meniscus region in most images than manual segmentation (in green). We continue to work on developing a fully automated meniscus segmentation algorithm.

Figure 1 : An example of a Sagittal view of a (a) normal knee, (b) one with OA OARSI 1, (c) OARSI 2, and (d) OARSI 3. Notice that as the OARSI score increases, meniscus degeneration becomes apparent along with osteophytes and cartilage damage.

Figure 2 : . A series showing the manual segmentations (green) and automatic segmentations (yellow) of the lateral meniscus from a patient with OA progression.