Automated Segmentation And Tortuosity Analysis Of Nerve Fibers In Corneal Confocal Microscopy

Corneal Confocal Microscopy(CCM)is a kind of noninvasive?real-time and high-precision microscope,which can acquire images of several structures of living corneal,including corneal nerve fibers.Some indicators of nerve fibers such as length,density,curvature and grade of tortuosity can be used to diagnose corneal diseases or injuries.However,it is a tedious task to obtain these information manually.Therefore,it is of great clinical significance to realize automatic analysis of nerve fibers in CCM images.Based on the above contents,this paper proposes an automatic analysis method of nerve fibers in CCM images,including segmentation and grading of tortuosity.Firstly,for the segmentation of corneal nerve fibers,this paper proposes an automatic segmentation method of nerve fibers with the conditional generative adversarial networks based on the improved U-Net.The improvement of U-Net includes two aspects.First is the improvement of the network structure.After the up-sampling operation of the decoding part of four layer U-Net,MSC(multi-scale separation and fusion)module is added to improve the feature extraction capability of the network.Second,the loss function is improved.The difference in nerve fiber length between the gold standard and the prediction map is combined with the Dice loss function to guide the network to pay more attention to the segmentation of fine fibers.Then the improved U-Net is used as the generator of conditional generative adversarial networks(cGAN)to segment the corneal nerve fibers.This method carries out a four-fold cross validation experiment on the nerve fiber data set of 8 cases of 90 CCM images provided by Zhongshan Ophthalmic Center.The average Dice coefficient of segmentation was 88.84%,indicating that this method can effectively realize the automatic segmentation of corneal nerve fibers in CCM imagesSecondly,aiming at grading of corneal nerve fibers,this paper proposes a method of automatic grading of nerve fibers in CCM images based on random forest.In the preprocessing step,the segmentation map of corneal nerve fibers and further skeleton map were obtained by using the automatic segmentation method proposed in this paper.In the feature extraction step,features including the calculation of tortuosity index under multi-scale spline fitting,Hog features,and the direction inconsistency among the nerve fibers are extracted.In the training and testing step of random forest,clustering algorithm is used to balance training samples,and the ReliefF feature selection algorithm replaces the random feature selection step of the random forest.The experimental datasets were obtained from 410 CCM image datasets published by Cixi Institute of Biomedical Engineering,Chinese Academy of Sciences,China and 30 CCM image datasets published by the University of Padova,Italy.We used five-fold cross validation in two datasets,and the average accuracy of automatic classification reached 83.1%and 90%respectively,indicating that this method can effectively achieve automatic classification of corneal nerve fiber tortuosity in CCM images.