Application Of Intelligence Algorithms In Visual Optics

Intelligence algorithms are widely used in ophthalmic optometry,but there are few techniques for detecting meibomian dysfunction in vivo confocal images,especially for diagnosing its disease types assisted by deep learning.Therefore,one application of this thesis is to provide a deep learning algorithm to distinguish the image features of meibomian gland dysfunction from those of acinar obstruction,gland atrophy or normal gland ducts.The turbidness of fundus refractive medium leads to the has poor resolution of optical coherence tomography,blurred image and serious scattering problem.However,the technology to improve the quality of fundus image is still not mature.Therefore,the other application of this thesis is to provide a technique for reconstruction of blurred image in optical coherence tomography.The main research contents,conclusions and innovations of in this thesis are as following:(1)For the experimental research,we need to master basic theory and classical neural network algorithm for deduction,thinking and elaboration.See Chapter 2 for the specific content.(2)In vivo confocal images marked by medical professionals in ophthalmology were used to train the three deep convolutional network models of Dense Net121,Dense Net169 and Dense Net201 respectively.The models were tested separately using an untrained test set of 1,663 images and compared with medical expert diagnoses.The specificity and sensitivity of the model were calculated to evaluate the performance of the model.We found that Dense Nets series network is very suitable for processing meibomian gland images,where the highest accuracy of Dense Net169 is more than 97%.The sensitivity and specificity to normal group were 94.5% and 92.6%,respectively.The sensitivity and specificity for blocking group were 88.8% and 95.4%,respectively.The sensitivity and specificity of atrophy group were 89.4% and 98.4%,respectively.In addition,by analyzing the model with confusion matrix and subject operating characteristic curves,we found that the diagnostic results of the model showed relatively high accuracy.This provides a new method for automatic detection and classification of meibomian gland dysfunction.Through further development,this model will become an effective tool for medical staff to assist in diagnosis,as detailed in Chapter 3.(3)The data from the recruited 210 subjects in between December 2018 and October2020 were collected in the experiment,which provides three kinds of innovation data,fuzzy simulation method respectively are filter attenuation method,frequency domain algorithm processing,pixel formula processing,through processing the image data to train built generated against the derivative network of the network.Finally,the peak signal-to-noise ratio and structural similarity were used to evaluate the model performance,and statistical data were analyzed by SPSS software to obtain the difference value between kruskal-Wallis detection data groups.By comparing the data set results,it is found that the structural similarity and peak signal-to-noise ratio of the model repaired image and label image in the data simulation group were greater than those of the fuzzy image and label image,i.e.,22.9016±0.1881>18.5778±0.4774;0.4834±0.0302>-0.0051±0.0184,the calculation results of the four test sets have certain statistical significance.Through analysis,it is found that the model has a significant effect on improving the image quality.Therefore,this study has increased the feasibility of generating adjudgment network for fuzzy image reconstruction,so as to achieve clinical application.See Chapter 4 for details.