Image processing on medical application - Automatic methods to calculate the area of articular cartilage on a magnetic resonance image

Abstract

Digital image processing plays a more and more important role in the medical diagnoses. Most widely used method for visualizing internal anatomical features of human body is magnetic resonance imaging (MRI). MRI can be used for the detection of osteoarthritis (OA) which affects the cartilage in the joints. Several techniques for the segmentation of the cartilage in MRI scans of the knee have recently been developed. One goal of segmentation is to automatically determine area and volume of the cartilage. Due to noise, however, none of these approaches is satisfactory in fully automated segmentation of articular cartilage. In our research, we attempt to study and develop new automatic methods to extract the cartilage from MRI knee scans. From cartilage images, cartilage area and volume are then computed. Three automatic methods are applied for cartilage extraction. The first method is bi-directional scanning segmentations method (BSSM), which is based on two basic properties of intensity value: discontinuity (edge detection algorithms) and similarity (thresholding algorithms). It is also based on statistical analysis (curve fitting algorithms and average weight calculation). The second method is neural network classifier method (NNCM), which is based on artificial neural network. For each pixel on an input image, through a neural network classifier, it is classified as a cartilage pixel if network output value is 1. Alternatively, it is classified as a background pixel if network output value is 0. The third method is active contour models method (ACMM), which uses an initial contour that approximates the boundary of a cartilage to find the “actual” boundary. This is an innovative method because we apply BSSM to define the initial contour. We also apply NNCM to compute the external energy in active contour models algorithms. Our three methods have succeeded in automatically extracting the cartilage from input image. The cartilage area and volume obtained from cartilage image, which is extracted by BSSM, NNCM, and ACMM, are highly correlation with the results obtained from reference cartilage image (correlation value in each case p ≈ 1, R2 ≈ 1). Thus, cartilage area and volume assessments are precise, reliable and acceptable. Among those methods, the ACMM provides the best results. Considering noise and complexity of the image, each method has both advantages and disadvantages. BSSM work well where there is significant high contrast between cartilage and background regions. It often fails where the contrast is low. NNCM can work well in low contrast. However, this method does not work accurately if the cartilage pixels have similar features of background pixels. Those pixels are classified as background pixels. As a result, this method reduces number of cartilage pixels. The third method, ACMM demonstrates higher accuracy in extracting cartilage from an input image. ACMM not only can take advantages of BSSM and NNCM but also can solve the problems existing in BSSM and NNCM. Therefore, results obtained from ACMM are the most precise and acceptable.