Low-dose SPECT Image Reconstruction Based On Neural Network

Single-photon emission computed tomography（SPECT）is a commonly used imaging technique in clinics.The principle is to inject the radioactive tracer into the patient,and gamma photons will be emitted during of the attenuation of radioactive tracer.The ph otons received by the detector around the human body are processed into projection data by a computer.Finally,the reconstructed SPECT image is reconstructed by the reconstruction algorithm,and the reconstructed SPECT image can assist doctors in clinical diagnosis.However,radioactive tracers have specific harm to patients,so reducing the dose of radioactive tracers is particularly important.In SPECT imaging,reducing the activity of radioactive tracers can achieve a low radiation dose,but it will lea d to a decrease in the projection data of sparse view angles or the number of photons in each view angle.Direct reconstruction of sparse projection data can lead to severe artifacts and noise in the reconstructed image,which affects the diagnostic accuracy of doctors.Therefore,reducing the dose of tracer used in SPECT imaging under the premise of ensuring imaging quality has become a hot research direction.In recent years,deep learning technology has been widely used in many fields because of its ability to learn data features from large amounts of data automatically.At present,the neural network is very bright in the fields of medical images,such as image reconstruction,image segmentation,lesion detection,and image registration.In this thesis,neural network technology is used to solve the problems of artifacts and noise in low-dose SPECT reconstruction to ensure that the quality of reconstructed images meets doctors’ standard of clinical diagnosis.The main contents are as follows:（1）In order to solve the problem of artifacts and noise after direct reconstruction of sparse angle projection data,a new network structure is proposed by combining a convolutional neural network and residual learning technology to learn the mapping relationship between sparse angle projection data and full angle projection data.Through the powerful feature learning ability of neural network,the missing projection data can be synthesized from sparse angle projection data.At the same time,the multi-level residual learning technology can play a denoising role in projection data processing to repair sparse angle projection data and improve the quality of reconstructed images.（2）Each row of two-dimensional projection data（sinogram）is composed of data collected by an angle of the detector in the SPECT imaging equipment.There is a specific temporal relationship between adjacent angles.Therefore,this thesis combines LSTM with the convolutional neural network U-Net,which is commonly used in the field of medical images,and proposes a new network structure.Based on the temporal relationship between the adjacent angles of the projection data,the sparse angle projection data is repaired,and the quality of the reconstructed image is further improved.In this thesis,two neural network models are proposed based on the characteristics of the projection data in SPECT reconstruction.They are the multi-level residual U-Net network model obtained by combining the convolutional neural network U-Net with the multi-level residual technology and the network model fused by the circular neural network LSTM and the convolutional neural network U-Net.These two network models are applied to the synthesis of sparse angle projection data,respectively.The experimental results show that the reconstructed projection data have a better improvement in the quality of the reconstructed image,which proves the effectiveness of the proposed method in the reconstruction of low-dose SPECT images.