Summary:
Arthritis is a pathological condition that occurs a very big range of people not only in the elderly but also in young people. An injury can be a cause of a bone marrow edema. Magnetic Resonance Imaging is used to diagnose these pathologies. Short Tau Inversion Recovery (STIR) sequences and Proton Density (PD) with fat saturation (FS) sequence, i.e. PD with selective spectral saturation of fat are included in a knee routine examination protocol. Each category in these sequences is represented by a high signal. The aim of this thesis is to create a computational system that will use digital image analysis techniques to measure the differentiation between normal MRI bone marrow erythropoiesis, arthritis and bone edema from injury. The methodology that will be used will initially include the manual selection of interest areas (ROIs). Then, from each ROI 12 color and texture features were extracted (e.g. 1st class, 2nd class statistics, etc.) and a feature selection process was apply to to isolate the features that provide differentiation between the three categories (normal marrow, arthritis, bone edema from injury). The methodology was tested on 92 cases (29 bone marrow erythropeisis, 31 injury, 38 osteoarthritis) from ANIMOUS KYANOS STAYROS. From the initial set of 12 features, the rank-features criterion was applied as a method of selecting features. The Cubic SVM classifier was used to classify the samples. In particular, with regard to the separation between normal and pathological (osteoarthritis and edema from injury), the accuracy rate reached 75.50% using the features standard deviation, energy, correlation range, homogeneity, contrast range, skewness, homogeneity range and contrast. For the discrimination, between normal marrow and edema from injury the best features were energy, correlation range, standard deviation, homogeneity, skewness, contrast range, contrast, and the highest accuracy rate was 68.30%. For the separation between normal marrow and osteoarthritis the features standard deviation, energy, correlation range, homogeneity, kurtosis, mean value reached the percentage of 70.10%. Finally, the best accuracy rate for the separation between the two categories of pathology was 65.20% using typical mean values, skweness, correlation, contrast.
Keywords:
Pattern recognition system, Medical Image Analysis, textural features, MRI, bone knee edema, arthritis, bone erythropoiesis.
