The Research Of Diagnosis Algorithms For Fungal Keratitis Based On Confocal Microscopic Images

Fungal keratitis is an inflammation of the cornea that results from infection by fungal organisms.The early symptoms of fungal keratitis are not obvious and are easily confused,the lack of early effective diagnosis and treatment often causes serious complications and even blindness.Most of the primary diagnostic methods of fungal keratitis are effective,but they all have deficiencies.Among the various diagnostic methods,confocal microscopy,which is a new type of noninvasive inspection instrument of living body,can obtain medical tomography images of living structures with high resolution.It has a high inspection acceptance and positive rate,but the diagnosis depends on the subjective judgment of the doctor.In this context,this paper focuses on the research of diagnosis algorithms for fungal keratitis based on confocal microscopic images and realizes automatic cornea images recognition with high diagnostic accuracy,which can provide doctors with accurate and reliable diagnostic information.How to distinguish the key structures in normal images and abnormal images is one of the biggest research difficulties in this paper.It can be seen that there are corneal nerve fibers and stromal layers in normal corneal images.In abnormal corneal images of fungal keratitis,the background is disorderly,and there are various kinds of fungal hyphae and spores.This study puts emphasis on the research of diagnosis algorithms for fungal keratitis based on traditional machine learning and deep learning.This paper studies and implements the existing methods,and according to the images characteristics,a new algorithm framework of "Data augmentation + Image fusion" is proposed on the basis of traditional methods.It has a targeted and comprehensive experimental performance.The main work and results of this paper contain:(1)This paper studies and implements diagnosis algorithms for fungal keratitis based on traditional machine learning.The algorithm framework of "Feature extraction+ Classification recognition" is determined at first.In feature extraction,GLCM,PCA,2DPCA,and MBP algorithms are introduced and applied,and a new feature extraction method ARBP,which is more robust and targeted for the images,is proposed based on MBP and AMBP.In classification recognition,KNN,LR,and SVM algorithms are introduced and applied.Experimental results show that traditional machine learning methods perform well in this project.ARBP is the best feature extraction method,SVM is the best classification recognition method,and ARBP+SVM achieves the highest diagnostic accuracy with 98.24%.However,because the imbalanced database,all of the algorithm frameworks have the disadvantage of being limited in specificity.(2)This paper studies and implements diagnosis algorithms for fungal keratitis based on CNN.The concepts of the revolutionary deep learning and CNN are summarized at first,and according to the images characteristics,classic CNN networks AlexNet,VGGNet,and GoogleNet are introduced and implemented in this paper.Experimental results show that the three CNN networks achieve a breakthrough than traditional machine learning,and they overcome the shortcomings of specific limitations.Among the three networks,22-layer GoogleNet has the best performance,achieving 99.73%diagnostic accuracy;the accuracy of 8-layer AlexNet is 99.35%,which is slightly weaker,but the model is more simple and the training speed is faster;the 16-layer VGG16 has a large amount of network parameters.With a small amount of data,the diagnostic accuracy is 99.14%and the performance is the weakest.(3)In order to pursue higher accuracy in the medical field,this paper proposes a new algorithm framework of "Data augmentation + Image fusion" to explore a more targeted and more excellent algorithm framework.Firstly,the normal image is augmented by flipping to solve the problem of limited and imbalanced database;Secondly,the SCS algorithm is proposed for image preprocessing to highlight the key structures in images and filter out irrelevant information;Thirdly,image fusion methods MF and HMF algorithms are implemented to fuse the SCS enhanced images with the original images,new algorithm frameworks and new databases are formed.Finally,the traditional machine learning algorithm framework and deep learning network are separately integrated into the new algorithm frameworks to perform experiments.Experimental results show that the two new algorithm frameworks have improved the performance of traditional machine learning algorithms,AlexNet and VGG16 algorithm frameworks with varying degrees,and the new HMF-based framework is even better."Data augmentation + HMF + AlexNet" achieves a perfect trade-off between diagnostic performance and computational complexity with the diagnostic accuracy of 99.95%,and it is the algorithm framework with the highest overall performance in this paper.