Study On Machine Learning Algorithms For Detection And Grading Of Cataract By Using Fundus Images

Eye is an important organ of the human body which is comprised of various components but not limited to the iris,retina,optic nerve,cornea,pupil,macula,and lens.There are various ophthalmic diseases that can affect vision,or cause blindness such as age-related macular degeneration,cataract,diabetic retinopathy,glaucoma,and trachoma,etc.According to the World Health Organization(WHO)report on vision,there are about 2.2 billion people who have vision impairment due to various ophthalmic diseases,and out of these at least 1 billion people can be prevented or yet to be addressed.Cataract is the main source of visual impairment which causes more than 50% of total blindness.Globally,at least 52.6 million people have visually impaired and 12.6 million are blind because of cataract.World report on vision reveals that almost 90% of vision impairment cases are from developing and under-developed countries where they lack basic healthcare facilities and the professional ophthalmologists are overburdened.Cataract is a progressive eye disease and tends to blindness if not diagnosed at early stages.The manual screening of cataract is a tedious approach,indorses extensive inconsistency among retinal experts,and the effectiveness of this approach is based on experts' subjectivity.Moreover,manual screening methods are unable to keep apace with demand because of the rapid prevalence of cataract all over the world.Thus,it is necessary to develop an automated method for the detection and grading of cataract and to assist the ophthalmologists.Most of the baseline studies employ slit-lamp images or ultrasonic images for the detection and grading of cataract.The main limitations to use these imaging modalities include lack of availability in remote areas,costly apparatus,and requires professional ophthalmologists.On the other hand,fundus images are widely accepted by researchers which are cost-effective and gives the bird's eye view in order to capture in-depth detail of the retina.The traditional machine learning methods perform cataract classification with the help of handcrafted features.The manual extraction of pertinent features from the fundus image is painstaking and time-intensive effort.To encounter these problems,this thesis is intended to create fast,accurate,and intuitive methods which could be implemented for early-stage cataract detection and grading.The stages include dataset acquisition,image preprocessing and data augmentation,feature extraction,and classification.First,a dataset is collected from the local hospital in China(Tongren Hospital),which is comprised of 8030 fundus images of various levels of cataract.Then,image preprocessing techniques such as image-resizing,green channel extraction,histogram equalization,contrast limited adaptive histogram equalization,top-bottom hat transformation,and non-local mean denoising are used to enhance image features.Next,a novel Gaussian scale space-based data augmentation technique is used to make a balanced dataset and to overcome data insufficiency problems.After that,the deep learning-based models are used to extract pertinent features directly from the fundus images.Finally,the cataract detection and grading is performed by using proposed models.The commonly employed performance criteria i.e.sensitivity,specificity,precision,and accuracy are used to measure the effectiveness of these methods.The main contributions of this dissertation are as follows:1.Performed cataract classification on a larger retinal dataset and employed various preprocessing techniques such as image resizing,green channel extraction,contrast limited adaptive histogram equalization,top-bottom hat transformation,non-local mean denoising,and image filtering to reduce noise and improve the quality of retinal images.2.Proposed novel data augmentation operations i.e.Gaussian scale-space theory(GST)and general augmentation techniques on each grade of cataract to deal with dataset unbalancing and annotated data insufficiency issues.The proposed model with an augmented dataset has achieved significant performance in cataract diagnosis.3.Developed retinal image analysis method for the identification of cataract severity by combining self-organizing map(SOM)and radial basis function neural network(RBF).The proposed SOM technique has been used for the appropriate region of interest selection followed by the RBF neural network for four-grade cataract classification.4.Proposed a novel ensemble of transfer learning-based models and machine learning models for the diagnosis of cataract and to achieve better diagnostic performance and to reduce computational time.The proposed scheme uses the combination of transfer learning models such as VGGNet,Alex Net,and Res Net for automatic feature extraction,and then support vector machine for classification.5.Proposed a two-stage novel framework namely CRNN by combining the convolutional neural network and recurrent neural network.The proposed CRNN model fuses the advantages of both these models to preserve the spatial correlations among patches.Moreover,the transfer learning models are employed to extract multilevel features and to investigate how accurately these algorithms perform cataract detection and grading.The outcomes of these studies suggest that the proposed methods have great potential to detect and grade early-stage cataract.Further,proposed methods are appropriated as an alleviating tool for cataract progression,and to assist the medical practitioners and ophthalmologists.Moreover,these models lead a new perspective to detect other retinal diseases in the future.