Deformable Image Registration Method For 3D-CT Lung Based On Deep Learning

Lung CT has the characteristics of fast scanning speed and high imaging quality,and is widely used in the diagnosis and treatment of lung diseases such as pneumonia,lung cancer,and pulmonary carbuncle.When doctors read CT images,they need to observe multiple CT images of the same patient to determine the location changes of the lesions.Medical image registration technology can align images,reduce image differences caused by imaging equipment,patient posture,etc.,so that doctors can observe the differences between different CT images.At the same time,registration technology is also used in medical fields such as surgical navigation and radiation therapy.The traditional image registration method has high accuracy,but low efficiency,and it takes several hours to register a pair of images,which limits its clinical application.In recent years,deep learning has played a major role in the field of medical images,and more and more researchers have begun to apply deep learning to the field of medical image registration.However,most of the deep learning registration methods proposed so far are suitable for organs with small deformation,and they are not effective for organs with large deformation such as the lung.The lung is a typical moving organ and is easily affected by respiration.The registration task is more difficult and challenging.This study,in cooperation with the Second Hospital of Jilin University,collected lung CT images of some patients in different periods,and proposed a deep learning model for lung registration.The input of the model is the fixed image to be registered and the moving image,and the output is a dense displacement field,which can be used to spatially transform the moving image to obtain the final warped image.The loss function consists of the similarity between the warped image and the fixed image plus the regularization constraint on the displacement field.The model does not require additional annotation information during training,and the entire registration process is unsupervised.The model is mainly composed of two parts.The first part is a feature pyramid network proposed for the large-scale deformation characteristics of the lungs.The network outputs two feature pyramids composed of feature maps of different sizes.The second part is the displacement field prediction network.The network generates a displacement field for each layer of feature maps and use these displacement fields to spatially transform feature maps of different scales which can enhance the network’s ability to learn transformations of different scales.The local cost volume is also introduced into the displacement field prediction network,which effectively improves the performance of the model.The registration model proposed in this paper is trained and tested on the public datasets Dir-Lab with expert annotation points and the datasets provided by the Second Hospital of Jilin University.The mean TRE value on the Dir-Lab test set is only 1.77 mm,which is significantly lower than other deep learning methods.The test results of the data set of the Second Hospital of Jilin University show that the network structure proposed in this paper can learn the realistic deformation while ensuring the accuracy and maintain the topological structure of the original image.The registration model proposed in this paper can complete the registration task within 1second,and the registration time is greatly shortened compared to the traditional method that requires iteration for each registration.