A deep learning approach for slice to volume biomedical image integration

Biomedical atlas images obtained from multiple sources need to be aligned and transformed into a single coordinate system so as to be able to integrate and relate these different sets of data. Formally known as image registration, this process of image pre-processing has proven to be integral in a wide array of computer vision applications, most notably in the area of medical imaging. During the last decade slice-to-volume registration, a particular case of image registration problem, has received further attention in the medical imaging community due to the emergence of several medical applications of slice-to-volume mapping. This paper proposes a Convolutional Neural Network (CNN) based deep learning approach for registering a 2D image slice to the 3D volume of images in a Biomedical atlas. The proposed CNN model is trained to determine the distance and pitch values that are used to describe the position of the 2D slice in the atlas coordinate system. High-level features are automatically extracted from the training dataset of images, which addresses the limitation of shallow machine learning techniques for handcrafted features followed by the classification task. Then on the basis of predicted values of distance and pitch, the target image is registered to the 3D volume of images. Experimental results showing the effect on the similarity of images with variation in distance and the impact of varying the distances among the classes on the regression are presented. It was observed that using the successive images at a distance of 10 lead to the maximum accuracy. These results demonstrate the applicability of the proposed approach to slice-to-volume image registration.