Prior distance map for multiple abdominal organ segmentation

Given the significance of organ specificity among different patients, prior knowledge, including the shape of a single organ and the relative position of adjacent organs, is challenging to apply on multiple organ segmentation tasks. To overcome this limitation, this paper proposes a novel feature classification algorithm based on prior distance map (PDM) for multi-organ segmentation to increase the effectiveness of prior information. Distance conversion is performed on a gray scale image to obtain the PDM by using Manhattan distance conversion. Feature vectors, which are composed of BRIEF and Local Binary Patterns (LBP) features, are classified based on PDM by using random forest algorithm. Our algorithm is validated using the public dataset of MICCAI 2015 Challenge. Experimental results show that the proposed algorithm has improved the accuracy compared with the existing algorithms, reaching the accuracy rate (ACC) of 82.9% for the spleen, 77.4% for the left kidney, 89.1% for the liver, and 62.2% for the stomach.