Research On Medical Image Analysis Algorithms Incorporating Medical Knowledge

In recent years,data-driven algorithms have gradually become the mainstream of medical image analysis methods.However,current data-driven algorithms focus mainly on learning the mapping relationships between input data and output results,failing to effectively utilize highlevel knowledge information.In the field of medical image analysis,high-level knowledge includes the diagnostic experience of doctors,human organ shape knowledge,anatomical feature point position knowledge,etc.Medical knowledge has important reference value for image analysis,but it is difficult to be directly learned by current data-driven algorithms.This thesis improves the performance of medical imaging analysis algorithms by integrating medical knowledge,exploring three aspects of image segmentation,imaging diagnosis,and image registration,and preliminary achievements.The main works accomplished in this study are as follows:(1)Low-dose mice micro-CT image segmentation using anatomical knowledge.MicroCT images of mice are widely used in preclinical cancer research.Segmentation of major organs from mouse Micro-CT is a prior step to measure drug metabolism and localize tumor lesions.Most of the micro-CT images of mice use low X-ray dose,which causes blurring of soft tissue organ boundaries.It is difficult to achieve accurate segmentation by relying on the image information itself,and the knowledge of organ anatomy needs to be used to improve the segmentation accuracy.In this paper,a two-stage deeply supervised fully convolutional network is designed to incorporate shape information and location information of mouse organs as anatomical knowledge into a deep learning network.The average Dice for organ segmentation in mice in this study was 0.84,which was 18～26% higher than the Dice of the traditional method.(2)Non-small cell lung cancer PET/CT image diagnosis using doctors’ diagnostic knowledge.In recent years,deep learning diagnosis of lung cancer image is a research hotspot.Most existing algorithms learn diagnostic features from the image data but ignore the diagnosis knowledge of experienced doctors.In this paper,seven types of lymph node features summarized by the doctors are incorporated into the algorithm.A deep learning network fusing features of PET and CT is designed,and then the features of the doctor summarize are fused with image features by a fully connected layer.The sensitivity and Area Under the Receiver Operator Characteristic curve(AUC of ROC)of this method are 0.86 and 0.96 respectively,which are 13.8% higher than the sensitivity of human doctors.(3)MRI image registration incorporating anatomical landmark knowledge.During the follow-up of liver cancer surgery,it is necessary to align the pre-and post-operative MRI images in order to compare and observe the changes in the lesion region.Traditional image registration algorithms focus on optimizing the grayscale similarity measure between two images and cannot effectively align the anatomical feature points of interest to the doctor.In this chapter,the bifurcations of hepatic vessel points are added as a constraint to the robust point matching algorithm to achieve a mean landmark registration error of 1.02 mm,which is 13.51 mm less than the mean error of the traditional grayscale-based registration method.In this thesis,medical knowledge is integrated into the existing medical image analysis algorithm to effectively improve the performance of image segmentation,registration,and diagnosis.This study makes a preliminary exploration for the combination of medical knowledge and data-driven algorithm.