Blockwise Feature-Based Registration of Deformable Medical Images

Su Wai Tun, Takashi Komuro, Hajime Nagahara

Abstract: We propose a framework for feature-based registration of deformable medical images using a blockwise approach. In our approach, we apply an accelerated-KAZE (AKAZE) feature detector on the initial image frame and input image frames for feature detection, and the detected feature points in the initial image are divided into blocks based on their coordinates. Then, the best feature point in each block of the initial image is picked up by using the response values of the detected features. Tracking of feature points is performed by finding corresponding features between the blockwise features of the initial image frame and all the detected feature points of the current image frame. Our approach has good registration capability even on sparsely textured surfaces such as human organs, which allows our method to be applied for surgery assistance. We demonstrate the effectiveness of our approach using three stereo endoscopic videos.