Research On Eye Fundus Image Processing For Diabetic Retinopathy Diagnosis Using Deep Learning

Diabetic retinopathy(DR)is a complication of diabetes,which is the ultimate cause of blindness among many diabetic patients.It is a widely spread eye disease that impacts the world overall.The early diagnosis of DR during the large-scale of the diabetic population plays a vital role in controlling the prevalence of DR from severe vision loss.Currently,the analysis of medical images is employed to perform the diagnosis of DR.Digital fundus imaging(DFI)is one of the known medical imaging modality used for the early screening of DR by the computer-assisted diagnosis(CAD)systems.To develop CAD systems for DR in years of(2015-2020),many authors employed deep-learning(DL)based multi-layer architectures to get high accuracy.The usefulness of DL methods is at least three-fold to recognize DR.First,the DL method did not require feature engineering that is an essential but time-consuming task.Second,the DL algorithms are very fast to learn compelling features for classification tasks without domain-expert knowledge.Third,the DFI images contained many DR-related lesions,which are difficult to detect by image processing algorithms so that those small features could be easily identified by DL-based methods.Despite those facts,the DL algorithms also achieved promising results in different real-world applications.In the domain of DFI,the state-of-the-art deep learning-based methods got significant higher accuracy when DR was compared with non-DR two-category based decision.Accordingly,it is challenging to recognize five-category of severity-level of DR.To validate and train these DL networks;many training DFI images are required for previous systems to achieve up-to-the-mark results.Moreover,the DFI image is suffered by low-contrast,noise,and light illumination problems.Non-uniform color space is mainly utilized in those past systems.Hence,it is difficult for DL-based methods to recognize the severity-level of DR.To solve the problems as mentioned above,we developed an innovative solution in this dissertation for improving the performance of the severity level of diabetic retinopathy by using advanced deep-learning(DL)based methods.We proposed a new multi-layer based DL architecture known as active deep learning(ADL)to define fast training scheme and obtain informative features that help to get higher classification results.This ADL-based framework is compared with state-of-the-art systems on different DFI datasets,and the obtained results indicate that this new ADL-based approach is applicable in many real-time CAD systems to recognize the severity-level of DR.The main research contributions and innovations of this thesis are listed as follows.(1)We proposed a new DFI pre-process step to enhance the contrast and adjust light illumination problems in a perceptual-oriented color space.The image quality of DFI is improved by employing local contrast,color content,brightness,and texture enhancement techniques by using the combination of an enhanced version of CIECAM97s color appearance model along with wavelet-transform methods in a perceptual-oriented color space.It was observed by experiments that the DFI image artifacts such as low contrast,noise,non-uniform illumination decreased the performance of the DL-based classifier to extract useful features.(2)We provided a new segmentation-based method to extract optic disc(OD)and CUP regions from DF images to localize DR-related candidate regions easily.It is confirmed by state-of-the-art systems that the segmentation of optic disc(OD)and CUP played a vital role in the detection of other retinal anatomic features.Those features used in the past to the diagnosis of glaucoma and diabetic retinopathy(DR).As a result,this thesis proposed a new method that focused on the detection of inconspicuous DR components,which are not addressed by the current deep learning algorithms.The advantage of this segmentation and localization of OD and cup areas helps to localize the DR-candidate region easily.(3)We developed a new active deep-learning(ADL)based approach to achieve higher classification accuracy compared with other state-of-the-art severity-level recognition of DR eye-related disease.Current DR-based deep learning methods produced higher accuracy only for the recognition of DR and non-DR based binary decisions.Therefore,there is a dire need to apply DL algorithms to recognize five-level of severity DR.Also;the existing DR-based deep learning arts require a handsome amount of labeled images,which in practice is a time-consuming process.An active deep-learning-based approach was developed that does not require thousands of images to perform the training step.Since we have trained the ADL classifier through input patches,those patches were obtained from Kaggle and EyePacs data sources.Those DFI images were captured through a diverse range of lighting conditions,cameras,and capture positions.The researches on retinograph images based on preprocessing,segmentation of DR-related lesions and classification of five-stage of severity-level of DR are developed in a perceptually uniform color space to provide help to ophthalmologists to increase sensitivity and specificity of diagnosis DR in clinical practice,which ultimately provide the solution to develop early screening system of DR either for timely referrals or grading support.