A Convolutional Neural Network Approach for 2D/3D Medical Image Registration

We present a Convolutional Neural Network (CNN) regression based framework for 2-D/3-D medical image registration, which directly estimates the transformation parameters from image features extracted from the DRR and the X-ray images using learned hierarchical regressors. Our framework consists of learning and application stages. In the learning stage, CNN regressors are trained using supervised machine learning to reveal the correlation between the transformation parameters and the image features. In the application stage, CNN regressors are applied on extracted image features in a hierarchical manner to estimate the transformation parameters. Our experiment results demonstrate that the proposed method can achieve real-time 2-D/3-D registration with very high (i.e., sub-milliliter) accuracy.