The Research On TEM Image Segmentation Strategy Of Glomerular Basement Membrane

Chronic kidney disease has become an important disease threat to global public health,and renal biopsy is an important method in diagnosis of chronic kidney disease.By means of transmission electron microscopy(TEM),pathological changes of ultrastructure of glomerular cells can be observed to make further pathological diagnosis.The study pointed out that in the ultrastructure of glomerular cells,glomerular basement membrane(GBM)changes have close relationship with chronic kidney disease.For example,thin basement membrane nephropathy manifests that the GBM is getting thinner diffusely.Therefore,in pathological diagnosis,doctors often need to identify and measure the basement membrane.However,the texture of gray TEM image of GBM is complex,and there are many kinds of lesions,and the contrast between the basement membrane and the surrounding tissue structure is low.Therefore,using computer-added image processing technology to segment the GBM region,will be more rapid and intuitive to observe the morphology of the basement membrane,and it is conducive to the diagnosis of chronic kidney disease.In recent years,the research of image segmentation has been a hotspot in the field of image processing,and many segmentation algorithm were proposed,but the segmentation algorithm of glomerular basement membrane is only developed in recent 20 years.This is mainly because of the complexity of biological image,the process of sample preparation,low image contrast and the fuzzy structure characteristics.All of this greatly increased the difficulty and complexity of image analysis,so as to make the development of GBM segmentation algorithm limited in a certain extent.At present,the proposed method can be divided into two categories:semi-automatic segmentation and automatic segmentation.These methods are mainly based on gray feature,texture feature and gradient feature of the image,and are effective to those image which has a nice contrast or has a simple shape.But if the contrast of the image is low or the shape of the GBM is complex,the performance of the segmentation will be unstable.Therefore,the level of automation and the effect of segmentation still needs to be further improved.To resolve the existing problems,two methods are proposed to realize the automatic segmentation of the basement membrane.The first method is proposed to segment the glomerular basement membrane automatically by image patch matching.The algorithm can effectively search similarities between image patches,but due to the low contrast of the basement membrane and the complexity of the shape,it is difficult to get the best matching result if only searching from one reference image.Moreover,the efficiency will be low if the number of reference images is getting large.Therefore,the search range was extended from one reference image to multiple reference images,and an improved searching method was adopted to enhance matching efficiency.Then,the optimal patches were searching out and the corresponding label patches were extracted and weighted by matching similarity.Finally,the weighted label patches were rearranged as a new image of glomerular basement membrane,from which the final segmentation result can be obtained after morphological processing.Method two is an automatic segmentation of GBM based on random forest.The algorithm is based on bootstrap sampling technique,generating a new training sample set,then the decision tree is built for each bootstrap sample and these trees are combined as a forest.Then the new data will be put into the forest and the prediction can be made by voting.The random forest algorithm has high prediction accuracy and good tolerance to outliers and noises.However,because of the difference of gray scale between the test image and reference image,some pixels will be lead to a confused classification,resulting that the segmentation accuracy rate is not high.Then,the concept of multiple random forests was introduced.The number of forest was extended from only one to multiple,so that the pixels which have a close gray scale can be found,which can improve the segmentation accuracy of the glomerular basement membrane.On the glomerular TEM(transmission electron microscopy,TEM)dataset,the Jaccard coefficient of method one is between 83%and 95%,and method two is between 84.6%and 92%.Experimental results show the proposed two methods can achieve higher accuracy,which provides valuable information for pathological diagnosis of renal biopsy.