Research And Application Of Medical Image Segmentation Based On The Combination Of Prior Knowledge And Neural Network

Accurate medical image segmentation plays a critical role in the assistant diagnosis of some diseases and the research of human organ/tissue development pattern.Recent research demonstrates that some diseases would cause varying influence both on retinal layers and brain tissues and accurate segmentation results is a critical step for quantitative analysis.Therefore,how to obtain accurate segmentation and collaborative analysis has been attracted interesting of researchers.However,complicated tissues would be easily affected by the pathological changes,which leads to topological deformation.Additionally,low intensity contrast and weak tissue boundary of images can also bring difficulties for medical image segmentation algorithms.Fortunately,the combination of prior knowledge is helpful to improve the automatic organ/tissue segmentation results of medical images.Therefore,how to appropriately combine prior knowledge with neural network to obtain accurate results is worth studying.To illustrate how to automatically obtain medical image segmentation,we focus on construct prior knowledge guidance for different segmentation targets and apply them on tasks of retail layer segmentation and brain tissue segmentation.The main distributions of this paper are summarized as follows,(1)An adaptive guided-coupling-probability level set method combined with retinal prior knowledge is proposed to automatically segment retinal layer segmentation in uneven retinal images.We first leverage a convolutional neural network to generate initial coarse layer segmentation.Then,by combing fixed retinal layer sequence and adaptive layer thickness constraint,we construct a coupling probability level set function to obtain results.To better deal with different retail images,we further proposed to design a segmentation flowchart.By leveraging features of intensity similarity and layer thickness variation,we first detect fluids for each layer,then automatically choose global or local intensity fitting term to characterize layers.Experimental results demonstrate the proposed method can obtain accurate retinal layer segmentation for normal and abnormal eyes.(2)A multi-scale self-supervised learning framework combined with brain prior knowledge is proposed to accurately segment multi-site pediatric images with artifacts.In the training stage,we first present to train a segmentation model using downsampling images,to alleviate the influence caused by artifacts.And training a segmentation model to obtain tissue probability maps from original image space at the same time.Then,constructing a global anatomical guidance by upsampling results to the original image space.Finally,the tissue probability maps with global anatomical guidance are as inputs to train a final segmentation model.In the testing stage,we propose a self-supervised learning strategy to train a site-specific segmentation model based on a set of reliable training samples.The experimental results show that the proposed method can effectively alleviate the influence caused by artifacts to obtain accurate tissue results.