An iterative two-step method using thresholding and SVM to segment bones from knee magnetic resonance images

Abstract

Automatic and accurate bone segmentation has important medical applications. Thresholding-based segmentation is the most widely used method to segment the object of interest from the background. Although bone tissue is among the brightest tissues in MRI T2 images, bone has a similar intensity and comparable characteristics to particular other tissues, such as fat, which may cause misclassifications and undesirable results. We have proposed an automatic, accurate, and rapid, with less computational complexity and time segmentation method for the knee bone using iterative thresholding and Support Vector Machines (SVMs). The initial threshold value is first obtained by Otsu Thresholding to partition the image into two classes: bone and non-bone candidate areas. The SVM detected the bone region from the bone candidate areas based on location and shape. The iterative process significantly improved the thresholding value until the bone was identified. The post-processing step utilized a Canny edge filter and image opening to eliminate the undesired area and to more accurately extract the bone. The proposed segmentation technique distinguished between bone and similar structures, such as fat. The object (bone) detection rate was 1, and the average segmentation accuracy was 0.96 using the Dice Similarity Index.