Shoulder Lesion Classification Using Shape and Texture Features via Composite Kernel

Abstract

Axial proton density (PD) weighted magnetic resonance (MR) images of shoulder which has the ability to represent bone edema while preserving anatomical details, provides valuable information for the evaluation of traumatized shoulder. The low signal to noise ratio of PD weighted slices of MRI while being a powerful tool for the detection of the pathological conditions, can hamper the determination of the anatomical structures and has a negative effect on the classification success. This study focuses on the classification of pathologies of the humeral head resulting from trauma or instability. In order to diagnose the bone edema and structural changes of the humeral head by using images of low signal to noise ratio, the shape and texture information were used together and their contribution to the classification success was evaluated. The texture information was obtained from the gray-level co-occurrence matrix (GLCM) algorithm and shape information obtained from the pyramid of histogram of gradients (PHOG) algorithm were joined together by concatenation and composite kernel. The feature vectors obtained from experimental studies were utilized for classification purposes by support vector machines (SVM) and extreme learning machines (ELM) methods; the results were presented comparatively.