Single shot 2D3D image regisraton

Abstract

In this paper, we present a very deep, 11 weights layers, Convolutional Neural Network (CNN) regression model for single shot and real-time 2D/3D registration. Different from optimization-based methods, which iteratively optimize the transformation parameters over a scalar-valued metric function representing the quality of the registration, the proposed method exploits the information embedded in the appearances of the Digitally Reconstructed Radiograph(DRR) and X-ray images and employs CNN regressors to directly estimate the transformation parameters. Unlike previous CNN approach which adopts an indirect way to cast the original complicated problem as several parts, we train a much deeper network to handle this registration problem by continuing to endeep the Convolutional Neural Network. To fit zooming in and out of DRRs more effectively we further more design a multi scale convolution kernel network. Our experiment results demonstrate the advantage of the proposed method in computational efficiency and accuracy. The research may indicate that powerful Convolutional Neural Network can learn the highly complex regression function that mapping the raw image data to the registration parameters thus achieve high accuracy and real-time in 2D/3D registration in a direct way.