Automatic Knee Cartilage Segmentation in Magnetic Resonance Images
Osteoarthritis is a degenerative disease of the articular cartilage. In the knee, it has a great impact in the patient’s quality of life and is one of the main causes of total knee replacement. Advances in medical imaging techniques have led to the development of cartilage-morphology-based analysis techniques for the assessment of pathological or potentially pathological joints, but this requires precise image segmentation. The main objective of the present work is to design and implement the required algorithms for automatic segmentation of the articular cartilage in knee MRI’s.
An image set consisting of 176 MRI scans was used. The images were obtained in a 3T machine with a DESS (Dual Echo Steady State) sequence. Anisotropic diffusion filtering was employed for noise reduction and a contrast enhancement ad-hoc transformation was developed. A set of 24 features were extracted related to voxel intensity, spatial location, and local geometry descriptors such as the Hessian matrix and the structure tensor. A two-stage classification strategy was adopted using neural networks: first, cartilage vs background; and second, subdividing the cartilage between the femoral, tibial and patellar compartments.
The final classification stage achieved Dice similarity coefficients (DSC) of 0,666±0,042, 0,674±0.039 y 0,550±0,055 for femoral, tibial, and patellar cartilage respectively. In a morphological comparison analysis between the manual and automatic segmentations, good geometrical correspondence was observed, although some areas of over segmentation were detected, leading to possible improvements.