Design of Medical Image Registration System Based on Deep Learning

Abstract

Medical image registration is an important task and technical difficulty in the field of medical image processing. It is of great significance for clinical work such as image fusion and tumor growth detection. Image registration aims to find the spatial transformation that maps one image to another. The traditional registration method iteratively optimizes the objective function of each pair of images to solve the spatial transformation, which has the problems of long registration time and large amount of calculation. In recent years, with the wide application of deep learning in the field of medical image research, image registration based on deep learning has become a promising research direction. Although the supervised registration method based on deep learning has improved in registration speed and accuracy, it is still difficult to obtain the label information required by supervised learning, which promotes the development of unsupervised registration method based on deep learning. In this paper, the unsupervised registration model is embedded in the lung segmentation network. Through the lung boundary information obtained from the lung segmentation network, the smooth constraint and the non smooth constraint are weighted and fused to construct a spatial adaptive regularization constraint. The results show that the spatial adaptive regularization term can flexibly and flexibly constrain the learning of lung registration model, and can simultaneously achieve accurate registration of slip and smooth motion in lung 4DCT images.