Glomerular Microscopic Medical Image Segmentation Based On Convolutional Neural Network

Accurate glomerular microscopic medical image segmentation is one of the keys to reliable disease diagnosis in renal pathology.At present,most of the glomerular image segmentation research work uses traditional pattern recognition or machine learning methods.These methods need to design corresponding features for different images,and the analysis methods are complicated.In recent years,deep learning technology based on convolutional neural networks has greatly promoted the development of automated analysis of digital pathology,and has been widely applied in the field of medical image processing.In this thesis,the deep learning method based on convolutional neural network is applied to glomerular microscopic medical image segmentation.The Cascade Mask R-CNN Scoring glomerular segmentation algorithm based on the Improved Mask R-CNN and Cascade R-CNN is proposed.The main work is as follows:Firstly,we construct a glomerular microscopic medical image dataset with pixel-level annotation.For the first time,the improved Mask R-CNN segmentation algorithm based on convolutional neural network is applied to glomerular microscopic medical image segmentation.We scale down the anchor frame length in the region proposal network of Mask R-CNN algorithm and increase the number of deconvolution layers of the head mask branch in the overall network structure of the algorithm,which can reduce the omission of small target glomerulus detection and refine segmentation results.Secondly,the Io U threshold used for detecting positive and negative samples in the network is too low,which produces many false positive detection bounding boxes.In this thesis,the bounding box branch and mask branch of the Improved Mask R-CNN algorithm are interleaved and cascaded,and a new Cascaded Mask R-CNN algorithm is designed to realize multi-level optimization of segmentation results.Finally,in view of the inaccurate mask quality score of the current instance segmentation algorithm,we add a mask scoring module to the Cascaded Mask R-CNN algorithm network to correct the deviation between the actual mask quality and the mask score.The experimental results show that the Cascading Mask R-CNN Scoring glomerular segmentation algorithm designed in this thesis can display more accurate mask integrity score and achieve more accurate glomerular microscopic image segmentation.