Biomedical CT Image Retrieval Using 3D Local Oriented Zigzag Fused Pattern

In this paper, we introduce a new feature descriptor 3D local oriented zigzag fused pattern (3D-LOZFP) for retrieval of medical CT images. The existing local patterns such as local binary pattern (LBP), local tetra pattern (LTrP) etc. captures the relationship between the reference and its surrounding pixels in a circular fashion in a 2D plane. The proposed descriptor encodes the relation between the reference pixel and its neighboring pixels using three unique 3D zigzag patterns in four different directions in a 3D plane. Therefore a total of 12 effective 3D zigzag patterns are introduced to capture the relationship between the reference and its neighbors in a 3D plane. The 3D plane is constructed by passing the input image through a Gaussian filter bank producing multiple filtered images containing multi-scale information. The feature dimensions are reduced using quantization and a fusion based scheme. The retrieval performance of the proposed descriptor is investigated by conducting experiments on two benchmark CT image datasets and then compared it with several recent techniques. The experimental results in terms of average retrieval precision (ARP) and average retrieval recall (ARR) across two databases validate the retrieval supremacy of the proposed descriptor over other techniques in CT image retrieval.