A multimodal machine-learning graph-based approach for segmenting glaucomatous optic nerve head structures from SD-OCT volumes and fundus photographs

Abstract

Glaucoma is the second leading cause of blindness worldwide. The clinical standard for monitoring the functional deficits in the retina that are caused by glaucoma is the visual field test. In addition to monitoring the functional loss, evaluating disease-related structural changes in the human retina also helps with diagnosis and management of this progressive disease. The characteristic changes of retinal structures such as the optic nerve head (ONH) are monitored utilizing imaging modalities such as color (stereo) fundus photography and, more recently, spectral-domain optical coherence tomography (SD-OCT). With the inherent subjectivity and time required for manually segmenting retinal structures, there has been a great interest in automated approaches. Since both fundus and SD-OCT images are often acquired for the assessment of glaucoma, automated segmentation approaches can benefit from combining multimodal complementary information from both sources.The goal of the current work is to automatically segment the retinal structures and extract the proper parameters of the optic nerve head related to the diagnosis and management of glaucoma. The structural parameters include the cup-to-disc ratio (CDR) which is a 2D parameter and is obtainable from both fundus and SD-OCT modalities. Bruch's membrane opening-minimum rim width (BMO-MRW) is a recently introduced 3D structural parameter that is obtainable from the SD-OCT modality only. We propose to use the complementary information from both fundus and SD-OCT modalities in order to enhance the segmentation of structures of interest.In order to enable combining information from different modalities, a feature-based registration method is proposed for aligning the fundus and OCT images. In addition, our goal is to incorporate the machine-learning techniques into the graph-theoretic approach that is used for segmenting the structures of interest.Thus, the major contributions of this work include: 1) use of complementary information from SD-OCT and fundus images for segmenting the optic disc and cup boundaries in both modalities, 2) identifying the extent that accounting for the presv ence of externally oblique border tissue and retinal vessels in rim-width-based parameters affects structure-structure correlations, 3) designing a feature-based registration approach for registering multimodal images of the retina, and 4) developing a multimodal graph-based approach to segment the optic nerve head (ONH) structures such as internal limiting membrane (ILM) surface and Bruch's membrane surface's opening.vi

A 3D Deep Learning System for Detecting Referable Glaucoma Using Full OCT Macular Cube Scans

“…Moreover, subtle pathologic changes are difficult to detect using predefined sectors because all the data in each sector is summarized by a single index, which is not a sensitive method to assess early disease damage. 9 One of the important benefits of 3D volumes is the 3D spatial contextual information available, which can be a tremendous help in disease characterization that are ambiguous in an individual 2D B-scan 11 and, thus, algorithms may find important patterns that humans may not see. Whereas the normative database for the present Cirrus SD-OCT algorithm consists of 284 healthy individuals, 7 deep learning algorithms applied to the entire cube scan can learn over thousands (or even millions of cubes if available) to overcome poor representation of a small normative database.…”