Lung CT Image Registration Based On Convolutional Neural Network

Medical image registration is an important part of medical image analysis.The registration algorithm for lung CT images is helpful to discover the changing rule of lung tissue with respiratory movement from lung CT images.In recent years,the deep learning method has obtained superior performance on many computer vision problems of medical images.However,the deep learning method has not been widely introduced into the task of CT image registration.It is very promising to apply deep learning to lung CT image registration.This paper focuses on a convolutional-neural-network-based registration method for lung CT images.The dense optical flow field is predicted by a convolution neural network,and then the floating image is transformed by spatial transformation layer using the learned displacement.The objective function of the neural network consists of two terms.One is the dissimilarity between the fixed image and transformed image,the other is the regularization term of the displacement field.It is an unsupervised machine learning method,of which the advantage is that it does not require any additional supervisory information,which saves manpower and material resources effectively.A convolutional neural network structure is proposed to learn the displacement field at first.This architecture consists of an encoder to extract features and a decoder for precise localization,where the add operation is utilized to fuse the features from the encoder.Based on this architecture,the attention gate in the attention mechanism is introduced to study the role of convolution structure in medical image registration.Attention gates can suppress irrelevant areas in the images,highlight significant parts that are useful for specific tasks.To evaluate the performance of our proposed method,experiments are conducted on three public datasets of Lung CT images.The registration performance is evaluated from three measures,which are accuracy,visual effect and registration speed,respectively.Experimental results show that the proposed registration methods have higher accuracy and robustness than other deep-learning-based methods.The visual results show that the key points in different images can match well.In addition,the running time is hundreds of times faster than that of traditional methods.