CNN Based Grading and 3D Presentation of Maculopathy from Digital Retinal Images

Abstract

Macula is the most vital part of retina where the central vision is formed and any damage to macula could result in severe visual impairment or even blindness. The group of diseases that a ects macula are collectively known as maculopathy and the symptoms of maculopathy usually appear in late stages when it becomes very di cult to completely recover the subject's lost vision. There are many retinal imaging techniques which are used to visualize hu- man retina but optical coherence tomography (OCT) is the most widely used technique nowadays because it can show early symptoms of maculopathy by capturing retinal cross-sectional regions. Many researchers have worked on extracting retinal information from OCT images. However, to the best of our knowledge, there is no literature available that provides a complete suite for the extraction and identi cation of retinal layers along with the uid segments for the diagnosis as well as grading of maculopathy as per clinical standards. This thesis presents a robust framework that rst extract and characterize up to nine retinal layers along with retinal uids from OCT vol- umetric scans irrespective of their quality or acquisition machinery. Then, it utilizes the extracted retinal information for the diagnosis and grading of maculopathy. Furthermore, the proposed framework uses the segmented layers for the reconstruction of 3D retinal surfaces as well as for the 3D mod- eling of human retina. To extract retinal layers, the novel structure tensor graph search (STGS) framework has been proposed. STGS rst computes coherent tensors which highlights the layer variations and then using those variations, it traces the layers iteratively by decomposing a tensor with max- imum coherency into an undirected graph. After extracting the layers, the retinal uids are automatically extracted through the proposed TU-Net ar- chitecture. TU-Net is a hybrid architecture consisting of three convolutional neural networks namely TU-Net-1, TU-Net-2 and TU-Net-3. TU-Net-1 ex- tracts retinal uids from the candidate scan through semantic segmentation, TU-Net-2 takes the extracted uid map and identify intra-retinal and sub- retinal uids along with measuring their respective volume. TU-Net-3 is responsible for diagnosing and grading maculopathy as per the clinical stan- dards. Furthermore, the proposed framework utilizes the extracted layers for generating a highly detailed 3D presentation of retina through Bowyer- Watson based Delaunay triangulation algorithm. The proposed framework has been validated on publicly available Duke datasets (containing cumula- tive of 42,281 scans from 439 subjects), Biomedical Image and Signal Analysis ii dataset (containing 4,260 scans of 51 subjects), Zhang dataset (containing cumulative of 109,309 OCT scans) and local Amanat dataset (containing 372 scans of 9 subjects). The proposed framework achieved the mean accuracy of up to 94.62% for accurately extracting nine retinal layers, achieved the mean dice coe cient of 0.906 for accurately extracting the retinal uids, achieved the accuracy of 98.75% for correctly identifying intra-retinal and sub-retinal

uids and achieved the accuracy of up to 93.42% for grading maculopathy as per clinical standards. Moreover, the proposed framework has been compared with other state of the art solutions on di erent publicly available datasets where it signi cantly outperformed them in extracting retinal layers, retinal

uids as well as in diagnosing maculopathy.