Development of automated 3D knee bone segmentation with inhomogeneity correction for deformable approach in magnetic resonance imaging

Osteoarthritis(OA) analysis is one of essential task in health issues. 3D Magnetic Resonance Imaging (MRI) segmentation plays an important role in a highly accurate knee osteoarthritis diagnosis. 3D segmentation knee MRI is challenging task because of complex knee structure, low contrast, noise, and bias field inherent in MRI. Deformable model is one of the most intensively model-based approaches for computer-aided medical image analysis. However, most of deformable models require prior shape and training processing for segmentation [1]. In this paper, we propose a deformable model-based approach with automatic initial point selection to segment knee bones from 3D MRI containing intensity inhomogeneity. This approach does not require manual initial point selection and training phase so that large amount of human resource and time can be saved. Preprocessing performs inhomogeneity correction and extracts voxels of interest in order to prevent leakage the boundary of target objective. The proposed deformable approach is devised by modifying boundary information of a hybrid deformable model [2] to morphological operation. Automated selection of initial point is motivated by 3D multi-edge overlapping technique in the [3] method. Experimental results are demonstrated 3D model comparing with other recent methods of knee bone segmentation [27,28] and 2D slices on both synthetic image with inhomogeneity correction or not. Our approach compared against a hand-segmented ground truth from experts, we achieved an average dice similarity coefficient of 0.951, sensitivity of 0.927, specificity of 0.999, average symmetric surface distance of 1.16 mm, and root mean square symmetric surface of 2.01mm. The result shows that our proposed approach is useful performing simple and accurate bone segmentation for diagnosis.