Volumetric local directional triplet patterns for biomedical image retrieval

Image retrieval became an active and fast advancing research area since the last two decades. Volumetric Local Directional Triplet Pattern (VLDTP) for biomedical image retrieval is presented in the proposed approach. The existing spatial field technique encodes the spatial relationship in 2D plane by taking gray value difference between center pixels and surrounding neighbors. Whereas proposed approach, encodes the relationship between center pixel and its neighborhood pixel in 3D plane with four individual directions (0°, 45°, 90° and 135°). 3D volume is obtained using multi-resolution Gaussian filter bank from 2D image with different standard deviation. To observe the effectiveness of proposed algorithm, experimentation is carried out on three state of art biomedical database which are OASIS-MRI (MRI images), MESSIDOR (Retinal images) and NEMA-CT (CT images) databases. The retrieval accuracy of proposed algorithm is measured by using minimum distance measure and two-layer feed-forward neural network (ANN). The achieved retrieval accuracy of proposed system in terms of Average Retrieval Precision (ARP) is 59.85%, 55.76% 95.21% using minimum distance measure and 92.46%, 62.45%, 99.32% using ANN in respectively. From experimental result, it is demonstrated that proposed algorithm provide us significant progress in ARP than existing feature descriptor.