Research On Medical Image Segmentation Method Based On Image Prior Knowledge And Deep Learning

As a key national strategic project at present,smart medical treatment requires interdisciplinary integration and innovation of medicine,mathematics,artificial intelligence and other disciplines.Intelligent diagnosis of medical image is an important link in intelligent medical treatment,which can improve medical efficiency.As the premise of intelligent diagnosis,the main task of medical image segmentation is to extract the key areas that doctors are interested in,such as diseased areas and organs.These have a wide range of clinical value,such as quantitative and qualitative indicators of organs and diseased areas,which will help to monitor the development of the disease and aid treatment.Meanwhile,medical image segmentation is also an important prerequisite for 3D reconstruction and registration.The objects to be segmented in medical images usually have specific prior knowledge.How to use the prior knowledge to improve the segmentation effect is a hot and difficult issue in the current research on medical image segmentation processing.This paper studies the medical image segmentation method based on image prior knowledge and deep learning.The main research contents and innovations are as follows:(1)The existing level set regression methods obtain the level set function through network prediction,and converts the level set function into the segmentation result directly through the Heaviside function,without considering the prior knowledge of the target,and the segmentation results are prone to isolated outliers and low edge segmentation accuracy.In this paper,the variational structure of Heaviside function in original level set regression network is studied and the regular term of geodesic active contour is added.At the same time,a convex relaxation model of the regularized Heaviside function model is constructed,and the equivalence analysis between the solution of the convex relaxation model and the original model is given.Furthermore,the iterative convolution soft threshold module algorithm is designed to solve the variational model after relaxation,and the stability analysis of the algorithm is given.(2)The edge indicator function in the regular term of the traditional active geodesic contour is difficult to accurately locate the edge of the blurred object in the medical image.In this paper,an edge predictor module is constructed,which is converted into an edge indicator function and incorporated into geodesic active contour regularization.A medical image segmentation network model based on edge predicator and geodesic active contour regularization is proposed,and a differentiable iterative convolution soft threshold module is designed to solve the network model.(3)Medical image annotation is expensive,and most of the acquired medical images are unlabeled and only a few of them are labeled.Based on the existing dual-task consistent semi-supervised segmentation model,this paper constructs a dual-task consistent loss between the regularized Heaviside function and the segmentation results,and proposes a semi-supervised segmentation network model based on geodesic active contour regularization.The iterative convolution soft threshold algorithm is further designed to achieve semi-supervised segmentation of medical images.Numerical experimental results show that the proposed semi-supervised segmentation model and algorithm have higher segmentation accuracy and fewer outliers in the segmentation results.