Novel Algorithms For Glaucoma Diagnosis And Fundus Image Segmentation

Medical imaging technologies have been advancing over the past few decades.An image of human eye fundus captured through non-invasive method can provide sufficient knowledge about the overall health of the eye as well as early signs of many eyes related disorders like glaucoma.However,a manual checkup of patient’s fundus image can take up to hours.Reducing the total amount of patients that can be served in a single day.With the help of AI and deep learning methods,this process can be sped up and total time can be squeezed into few minutes.There have been many instances of works related to automated diagnostics of the eye diseases using AI based approaches.Thus,AI can work along ophthalmologists and clinicians to help detect certain illnesses.However,AI based automated diagnostics approaches comes with their own set of challenges and short comings Therefore,this thesis works towards developing novel techniques for retinal health,fundus segmentation,and glaucoma detection.The project involves following pipeline for a robust glaucoma detection and segmentation:1.Segmentation of the Vasculature of the fundus images2.Segmentation of the Optic disk(OD)and Optic cup(OC)of the fundus images using inductive transfer learning to find the ratio between the OC and OD3.Classification of images into normal and abnormal categories using features extracted from the segmentation tasksThe thesis work proposes an automated glaucoma detection scheme that utilizes a novel Unet-based CNN architecture named DilUnet as a backbone.DilUnet architecture is combined with a unique and smart training strategy comprised of preprocessing pipeline,inductive transfer learning and image processing operations to complete the feature extraction framework.The extracted features are combined with a novel wavelet scattering transform and PCA based projection error to train machine learning classifiers for the glaucoma detection.For the task of fundus blood vessels segmentation,DilUnet is trained from scratch along with set of pre-processing steps and real time augmentation.This allows the model to learn features general features about the fundus images as well as other features that are unique to fundus blood vessel segmentation.The knowledge and features learned in the task of blood vessel segmentation are then utilize via an inductive transfer learning strategy to achieve the OC and OD segmentation.This strategy does not require the model to be trained from scratch once again.And since the task of both blood vessel segmentation and OC/OD segmentation are quite similar,the segmentation performance is not compromised.Both segmentation tasks play an important role in the detection of glaucoma since they are employed to extract glaucoma related clinical features i.e.,Blood vessels ratio(BVR),horizontal and vertical Cup to Disk ratio and ISNT rule-based features.Finally for the detection of glaucoma,the clinical features are combined with wavelet scattering based textural features to form the final feature set.These features are then reduced to a novel PCA based projection error.Projection error is fed as input to a machine learning classifier like SVM,random forest or decision tree for the ultimate detection of glaucoma.From the Novel CNN architecture for segmentation to the PCA based projection error,All the Objectives of this work are synergetic and once put together,provide powerful complementary support to each other.While at the same time techniques used in this work can fully stand on their own as individual components for other image processing-based tasks.Proposed DilUnet can be used in several different scenarios as a standalone all-purpose CNN architecture for segmentation,object detection and even as a generator.Compact scattering features and PCA based projection error can be utilized in fast paced machine learning tasks to provide robust classification.The evaluation metrics revealed that the segmentation performance of the DilUnet is higher than the original U-net and other U-net based architectures as well as many other state-of-the-art segmentation techniques and it is robust to noise.A breakdown of the proposed glaucoma detection method showed that it surpasses deep learning-based classification due to its conveniency and simplicity.And the demonstration of the novel projection error’s performance deemed it superior to other machine learning based techniquesThis work provides the ophthalmologists/clinicians with two options.1:They can either utilize the complete classification pipeline for the rapid automatic detection of glaucoma or 2:They can make use of DilUnet based segmented blood vessels,OC/OD and subsequent biomarkers computed through these segmentations to manually diagnose the glaucoma.With the aid of DilUnet segmentations,such manual detection of glaucoma is way more accurate and time efficient than the traditional ways of eye examining.Moreover,the results deduced from this kind of manual checkup can be aggregated with the automatic glaucoma detection to make the diagnosis procedure even more accurate.