Automatic segmentation and analysis of Magnetic Resonance images of the knee bones and cartilages

The aim of this research was the development and evaluation of a non-invasive toolto allow the automatic, accurate and reproducible quantitative measurement of subtlechanges in knee cartilage tissue over time and across a patient population. Non-invasiveinformation was obtained by acquiring, segmenting and analyzing high resolution MagneticResonance (MR) images of the whole knee joint.The interest in knee cartilage is from Osteoarthritis (OA), a disease characterizedby changes in structure and degeneration of cartilage tissue. The X-ray radiography isthe standard imaging modality for diagnosis. However, cartilage tissue is not directlyimaged so imprecise surrogate measures are used. It is now generally accepted thatradiographs do not provide the sensitivity required to perform short term studies or drugtrials into OA. Magnetic Resonance (MR) imaging allows the non-invasive acquisitionof high resolution images of the whole knee joint, including the cartilages. MR has thepotential to perform accurate and repeatable quantitative measurements of the cartilagesand allow the detection of small changes over time. This is important as it will aidthe development and refinement of cartilage-dedicated therapeutic strategies, surgicaltreatments and identification of promising drug targets.Quantitative measurements require the cartilage to be segmented, a task whoseaccuracy can significantly influence the error and reproducibility of the analysis. This iscritical as only small volume losses (≃ 5% a year) and sub-millimetre thickness changesare expected. Due to the structure and morphology of the cartilages as well as the natureof MR acquisition, obtaining accurate segmentations can be problematic and are usuallyperformed manually by trained professionals. The hypothesis investigated was thataccurate, reproducible quantitative measurements of cartilage tissue can be obtainedautomatically by developing and using advanced image processing techniques andmathematical models, particularly the use of trained models of shape and appearance.This was investigated using MR images from two sources 1) previous and currentstudies that were kindly provided by fellow researchers and collaborators and 2) acquisitionsof volunteers. The MR images from different sources were acquired using differentMR sequences, parameters and field strengths. The images had a mixed demographicof participants (male or female and young adult to elderly) who had healthy knee jointsor had mild to moderate OA. In some cases multiple scans were acquired from the samepatient, either on the same day or longitudinally to allow reproducibility and test-retestexperiments to be performed.A hybrid segmentation scheme was developed to process the MR images. This involvedthe initial estimation of the bone location, accurate segmentation of the bones,extraction of the bone-cartilage interface, estimation of cartilage properties and thensegmentation of the cartilages. The initial location of the bones was estimated using arobust affine registration to an atlas, the bones were segmented using 3D active shapemodels, while the bone-cartilage interface and cartilages were processed using priors,statistical thickness models, image information and other constraints. Three other approacheswere also investigated, 1) an intensity non-rigid registration based on free formdeformation modelled using B-splines, 2) an improved watershed algorithm, and 3) atissue classifier that utilized bone presegmentations. These approaches were developedand applied to the segmentation of individual magnitude MR images.The acquisition and use of additional MR information can improve the accuracy ofsegmentation algorithms. This was investigated using phase MR information, multipleMR sequences and multiple echos. This additional information was utilized by trainingclassifiers to segment bone tissue using features extracted from the magnitude and/orphase of the MR signal. This was extended to incorporate shape information to allowmore accurate and anatomically valid segmentations to be obtained.Manual segmentations performed by experts were used as the ground truth, withvoxel based and surface based measures used to evaluate the segmentation quality ofthe automatic segmentations. The Dice similarity coefficient (DSC) ranges from 0 to 1and indicates the spatial overlap. The hybrid segmentation scheme obtained an medianDSC of (0.89, 0.96, 0.96) and (0.833, 0.826, 0.848) for the (patella, tibia, femur) bonesand cartilages. This was significantly better than the (0.810, 0.793, 0.849) and (0.732,0.785, 0.758) obtained by the tissue classifier and non-rigid registration respectively. Theaverage DSC obtained for the whole cartilages by a semi-automatic watershed algorithm(0.896) is slightly higher (0.891), however the hybrid approach is completely automaticand obtains separate cartilage plates. Quantitative measures found a median volumedifference of (5.92, 4.65, 5.69)% and absolute thickness difference of (0.13, 0.24, 0.12) mmfor the (patellar, tibial, femoral) cartilages.The use of additional information, in particular phase and magnitude information,allowed more accurate bone classification results to be obtained with an average DSCof 0.907 for the bone. The use of shape information further improved the robustness,accuracy and anatomical validity of the segmentations, and obtained an overall DSC of0.922. Further investigation is required to determine whether additional informationcan improve cartilage segmentation.The innovative segmentation system developed allows the non-invasive automatic,accurate and reproducible quantitative analysis of healthy and diseased cartilage tissuefrom MR images. The presented scheme is competitive with semi-automatic methods,with only a slight loss in accuracy (volume CV ≈ 5 − 6% compared to ≈ 2 − 3%), whichis more than made up by it being fully automatic, requiring no user interaction besidesa cursory visual validation of results.