Research And Application Of Low Dose Cone Beam Computed Tomography Image Reconstruction Based On Deep Learning

Radiation therapy,along with surgery and internal medicine therapy,is one of the three major methods of tumor treatment.To ensure that the actual treatment process is consistent with the radiotherapy plan and make the radiation accurately strikes the tumor,it is necessary to use the Cone Beam Computed Tomography(CBCT)device onboard the medical linear accelerator to quickly obtain CBCT images of the patient’s treatment site before implementing radiotherapy.On the one hand,it is used to correct the patient’s posture,and on the other hand,it can be used to observe the changes in the patient’s tumor and normal tissue structure around the tumor,so as to timely revise the radiotherapy plan.Compared with CT images,CBCT has the characteristics of lower dose and more convenient access.At present,CBCT only has single level ray imaging,and the commonly used CBCT imaging energy is 80 ke V,which can avoid patients receiving additional radiation exposure.However,the organ tissue edges of the 80 ke VCBCT image are relatively blurry and cannot effectively suppress the generation of metal artifacts,making it limited in observing changes in the patient’s target area.The use of 140 ke V imaging can significantly suppress artifacts in the image,improve image quality,and thus improve radiotherapy accuracy,but at the same time,it increases the burden on patients.Therefore,we propose a deep learning method.By collecting low-dose 80 ke VCBCT images,we can reconstruct and generate images closer to 140 ke V energy level,reduce the radiation exposure received by patients,and at the same time provide radiotherapy technicians with higher quality CBCT images,so as to improve the observation effect of tumors and their surrounding normal tissues,and achieve the purpose of improving the accuracy of radiotherapy.The main work of this thesis includes the following aspects:(1)A dual energy CBCT image dataset was created,and through targeted data collection in Shandong Cancer Hospital,80 ke VCBCT and 140 ke VCBCT image data of the head and neck of fourteen patients and phantoms were obtained in pairs.3Dslicer was used to register140 ke VCBCT based on 80 ke VCBCT images,obtaining more reasonable label data,laying the foundation for future image reconstruction work.(2)A CBCT image reconstruction algorithm based on residual and attention mechanism of convolutional neural network is proposed,and the mixed loss function is used to guide the training network,which can suppress the noise and artifacts of the image while better preserving the edge and detail information of the image.Firstly,this algorithm better recovers image features through residual connections and attention mechanisms.Secondly,the algorithm fuses the Mean Absolute Error loss function with the Mean Structural Similarity Index loss function to focus on the similarity of image structure,so as to realize the reconstruction task of low-dose CBCT images.(3)A low-dose CBCT image reconstruction platform software system was designed based on specific practical needs.The system includes functions such as physician login,reading and visualizing patient CBCT images,and low-dose CBCT image reconstruction.This system can assist radiotherapy technicians in better correcting the registration results between CBCT and radiotherapy plan CT,and improve the accuracy of radiotherapy positioning.