Research On Automatic Grading Algorithm Of Cataract Fundus Image

In China,almost 90% of the elderly above 60 years of age with impaired vision have cataract diseases,and around 90% of the eye diseases are diagnosed by observing the fundus,so the observation of fundus images has always been an essential mean of diagnosing a cataract.However,it is highly uncertain about judging the degree of lesions based on experience,but also the efficiency of this method is very low.Therefore,employing a computer-aided diagnostic system to perform the automatic grading of cataract is of the excellent research value of practical use.Considering that the clarity of blood vessel network is an important indicator of the degree of cataracts when ophthalmologists diagnose cataracts by observing the fundus image,and they often focus on the blood vessel around the optic disc.Therefore,based on the idea of a blood vessel mask assists automatic grading of cataract,the performance of artificial(deep learning)feature based method in the task of computer-aided cataract diagnosis was explored.By designing the comparison experiment,the effectiveness of the two methods mentioned above was highlighted and signified.And also compares the applicability of these methods.The precise contents of this researh are as follows:(1)The method based on the artificial feature: The method based on artificial feature was adopted to realize the grading of cataract fundus images,in particular the method based on adaptive window model(AWM)matching was proposed to enhance the contrast of fundus image of blood vessel and the background,then the spoke feature of the cataract enhanced image was extracted to realize automatic grading of cataract fundus images.(2)The method based on the deep learning feature: The method based on deep learning feature was adopted to realize the grading of cataract fundus images,in particular the initial blood vessel mask generation model and the cataract fundus images grading model with blood vessel attention mechanism were co-optimized to improve the grading performance of the latter.Considering the availability of a labeled blood vessel dataset that can be used to pre-train the initial blood vessel mask generation model,an initial blood vessel mask generation model based on U-Net is first pre-trained.Later,the initial blood vessel mask of the data with cataract grade labeling was obtained;furthermore,it was refined through the attention mechanismbased on the global context vector generated by the automatic grading model of cataract fundus images.The refined blood vessel mask is not only used to guide the feature extraction of the automatic grading model of cataract fundus images.But it can also be used as semi-supervisory information that can further optimize the initial blood vessel mask generation model.(3)In the experimental phase,to explore the relationship between the performance of the method based on artificial(deep learning)feature and training set size,the dataset used in this research was expanded,and the grading of cataract fundus images was realized by the two methods mentioned above before and after the expansion.The experimental results show that the average classification accuracy of the method based on artificial(deep learning)feature reached up to 80.12%(88.57%),which was better than of existing methods;Moreover,the method based on artificial feature was better than of techniques based on deep learning feature before data expansion,but the method based on deep learning feature was far better than the method based on artificial feature after data expansion.