Research On Automatic Recognition And Classification Model Of Fungal Hyphae In IVCM Images Based On Deep Learning

Objective: The purpose of this study is to use a deep learning algorithm to construct an intelligent system that automatic recognition and classification of fungal hyphae form the in vivo confocal microscopy images.By comparing the performance of the model constructed by different training samples,the reasons that affect the performance of the model are analyzed and provides reference for further research on the intelligent platform to assist doctors in diagnosing fungal keratitis.Methods: There are 2623 images in the ophthalmic IVCM image library of our hospital(from April 2018 to November 2019)that met the standards were selected,which were divided into train set and test set.The train set included 2088 images : there are 688 images of fungal hyphae,368 images of inflammatory cells,303 images of activated dendritic cells,380 images of normal cornea,349 images of diabetic corneal neuropathy.Then to import those datas of this two groups into the Residual Network(Res Net-101)and obtain the model 1(Res Net1)and the model 2(Res Net2)which with automatic recognition and classification of IVCM images with the fungal hyphae.Through 535 pictures(172 with fungal hyphae and 363 with non-fungal hyphae)of the test set.The performance of the two models was evaluate.The area under curve(AUC),the accuracy,sensitivity and specificity were recorded in the external validation results,as an index to evaluate the value in clinical application of this Res Net-101.By comparing the indicators of the two models,we can judgment whether the effect on the construction of the model when to increase the pictures of diabetes.Results: Through the external verification,the Res Net1 can correctly recognize 515 images,misrecognize 20 images(including 6 false negative and 14 false positive).The AUC,accuracy,sensitivity and specificity were 0.986,0.963,0.919 and 0.983 respectively.There are 501 images were correctly identified and34 images were incorrectly identified(including 4 false negative and 30 false positive)in the Res Net2,The accuracy,sensitivity,specificity and AUC were0.936,0.826,0.989 and 0.978,respectively.When 349 IVCM images of diabetes were added to the train set,the accuracy,sensitivity,specificity and AUC of Res Net2 model decreased compared with Res Net1 model,especially the sensitivity.Conclusion: The intelligent system based on Res Net-101 has a high precision ratio in identifying fungal hyphae in IVCM images of fungal keratitis.However,the results of Res Net2 model decreased after adding the images of the diabetic corneal neuropathy which shows that increased this dataset has an impact on the results of this algorithm.