Multimodal Glaucoma Classification Using Segmentation Based Biomarker Extraction
Rahil Parikha, Van Nguyenb, Anita Penkovac
aDepartment of Computer Science, University of Southern California, Los Angeles, 90089, California, USA
bRoski Eye Institute, Keck School of Medicine, University of Southern California, Los Angeles, 90033, California, USA
cDepartment of Aerospace and Mechanical Engineering, University of Southern California, Los Angeles, 90089, California, USA
Email: {rahilpar, vann4675, penkova}@usc.edu
aDepartment of Computer Science, University of Southern California, Los Angeles, 90089, California, USA
bRoski Eye Institute, Keck School of Medicine, University of Southern California, Los Angeles, 90033, California, USA
cDepartment of Aerospace and Mechanical Engineering, University of Southern California, Los Angeles, 90089, California, USA
Email: {rahilpar, vann4675, penkova}@usc.edu
Abstract
Glaucoma is a progressive eye disease that leads to irreversible vision loss and, if not addressed promptly,
can result in blindness. While various treatment options such as eye drops, oral medications, and surgical
interventions exist, the disease may still progress. Therefore, early detection and diagnosis can mitigate
the devastating vision loss of glaucoma. The primary focus of the proposed study is the development of a
robust pipeline for the efficient extraction of quantifiable data from fundus images. What sets our approach
apart from other approaches is the emphasis on automated feature extraction of glaucoma biomarkers from
particular regions of interest (ROI) along with the utilization of a weighted average of image features and
clinical measurements. Our unique approach leverages segmentation based models in conjunction with
computer vision techniques to efficiently extract and compute clinical glaucoma biomarkers such as optic disc
area and optic cup to disc ratio. The extracted biomarkers are consistent with trends observed in clinical
practice, thus supporting the validity of the feature extraction approach. While subtle disease progression,
inconsistencies in image quality and a general lack of metadata impact classification performance, the
weighted combination of image based features with glaucoma biomarkers achieves a test accuracy of 91.38%
for glaucoma classification, successfully addressing the limitations of traditional single-modality approaches
such as fundus imaging and optical coherence tomography (OCT).