3-D automatic segmentation and modelling of cartilage comparments in high-field megnetic resonance images resonance images of the knee joint

Abstract

Magnetic resonance (MR) images enable morphological and quantitative assessment of cartilaginous anatomic structures through their manual or automatic segmentations. Because structural changes in the joint compartments, especially deterioration of cartilaginous tissues, indicate strong correlation with a disorder so-named as osteoarthritis, quantification and visualization of the cartilage can establish evidence or progression of this disorder as well as effectiveness of therapeutic or surgical practices. In this thesis, fully-automatic segmentation and modelling of the whole femoral cartilage (FC), tibial cartilage (TC), and patellar cartilage (PC) compartments in MR images of the knee joint was mainly aimed avoiding segmentation methods specialized for the anatomical structures of interest and considering systems with limited resources in particular. The secondary purpose of the thesis was to investigate if detection of the image features such as edges or interest points directly in three-dimensional (3-D) volumes, rather than in two-dimensional (2-D) slices as usual, brings about some advantages or not for volumetric images. In the first study presented in this thesis, all cartilaginous compartments in the knee joint were automatically segmented in high-field MR images obtained from Osteoarthritis Initiative using a voxel-classification-driven region-growing algorithm with sample-expand method. Computational complexity of the classification was alleviated via subsampling of the background voxels in the training MR images and selecting a small subset of significant features by taking into consideration systems with limited memory and processing power. Different subsampling techniques, which involve uniform, Gaussian, vicinity-correlated (VC) sparse, and VC dense subsampling, were used to generate four training models. The segmentation system was experimented using 10 training and 23 testing MR images, and the effects of different training models on segmentation accuracies were investigated. Experimental results showed that the highest mean Dice similarity coefficient (DSC) values for all compartments were obtained when the training models of novel VC sparse subsampling technique were used. Mean DSC values optimized with this technique were 82.6%, 83.1%, and 72.6% for FC, TC, and PC, respectively. This study did not require finding a volume of interest, segmenting a bone, or determining bone-cartilage interface prior to segmentation of a cartilage compartment unlike most of the related studies in the literature. Therefore, computational complexity of such a prior operation was reduced in the system. Also, despite processing MR images with single modality for only osteoarthritic participants, the system obtained accuracies similar to those of the related works. About 30-min processing time was promising for segmenting all compartments in all slices of an MR image on a resource-limited platform. Moreover, a novel hybrid segmentation method was proposed to primarily deal with the oversegmentation problems of the former system. This method combined the results of voxel classification-based segmentation with results of active appearance model (AAM) segmentation of the cartilage compartments through an information fusion procedure. Experimental results for only FC compartment using the same sets of training and testing MR images indicated that AAM segmentation could approximately determine the appearance information of the compartments in most of the testing MR images. However, failure in some of the MR images prevented implementation the information fusion module as intended. Simply intersecting the segmentation results of the tissue classification and appearance modelling modules for information fusion, the hybrid segmentation method could not outperform the former voxel classification-based segmentation method with its highest mean DSC value of 73.78% for FC. With regard to the secondary purpose, standard Marr-Hildreth edge detection and Harris corner detection methods were extended to run in 3-D volumetric images. The results of the standard methods, which were applied in 2-D slices of the volumetric images, were qualitatively compared with results of the 3-D methods. As a result, in knee MR images, 3-D Marr-Hildreth method prominently detected the principal bone and cartilage edges found by the standard 2-D Marr-Hildreth method gaining additional sensitivity to gradient changes along the slices. In volumetric images of FC, the proposed 3-D Harris corner detection method determined well-localized and more distinct interest points at salient positions close to the surface boundaries.