Multi-dose CT Denoising And Performance Assessment

Computed Tomography（CT）is one of the important auxiliary diagnostic methods in the modern medical system.High-quality CT images help doctors observe the patient’s lesions,thereby improving the accuracy of diagnosis.In order to reduce the damage of radiation to the human body,low-dose CT is now used.During the clinical diagnosis,the CT instrument automatically adjusts the radiation dose to produce a variety of different low-dose CTs.Low-dose CT introduces noise,so it is of great significance to study CT image denoising technology.In addition,the existing medical image quality assessment mostly adopts natural image assessment methods,which cannot accurately assess the important texture information in medical images.Therefore,it is necessary to study medical image quality assessment methods.For medical image quality assessment,this thesis proposes an image quality assessment framework combining target detection and radiomics;for one dose in multiple doses,this thesis integrates radiomics features to perform low-dose CT denoising;In order to solve the problem of poor generalization of multi-dose CT denoising,a multi-dose CT denoising model based on multi-task learning is proposed.The specific description of this work is as follows:（1）An image quality assessment framework combining object detection and radiomics is proposed for the first time.Aiming at the problem that the existing image quality assessment methods cannot accurately assess the quality of medical images,object detection and radiomics are studied in this thesis.By analyzing the object detection results and radiomics features of multi-current dose and multi-voltage dose CT images,the positive correlation between CT radiation dose and object detection results is proved.It is also demonstrated that radiomics features can characterize the effect of dose changes on image quality.Experiments show that the medical image quality assessment framework combining object detection and radiomics is effective.（2）Low-dose CT denoising combined with radiomics features is studied.The existing low-dose CT denoising methods has insufficient ability to preserve image texture details.In response to that problem,the influence of radiomics features on low-dose CT denoising is studied.Two loss functions based on radiomics are proposed.Objective assessment indicators,object detection and radiomics combined indicators are compared to evaluate the improvement.Through the results,it is proved that the improve method based on radiomics features can enhance the denoising ability and texture detail preservation ability of denoising networks.（3）A multi-dose CT denoising network based on multi-task learning is proposed.The existing low-dose CT denoising network has good noise reduction ability,but cannot guarantee the generalization performance of clinical multi-dose CT images.Aiming at the above problems,this thesis simulates four different radiation dose CT images to construct a multi-dose CT data training set,and designs a multi-dose denoising network.Experiments are carried out on the anthropomorphic phantom multi-dose CT dataset,and two indicators are used for evaluation: objective indicators,object detection and radiomics combined indicators.The results prove that the multi-task learning network proposed in this thesis can improve the denoising ability and have better generalization performance.