Abstract
Purpose :
Artificial intelligence (AI) has been used successfully to diagnose several eye conditions, including glaucoma. OCT-Angiography (OCTA) is a non-invasive technology useful for glaucoma diagnosis and follow-up. To our knowledge, no study has applied AI in OCTA scans to increase the diagnostic capacity in glaucoma. Our purpose was to develop an AI tool to assist in glaucoma diagnosis using OCTA scans. We also intended to measure its performance using the area under the curve (AUROC), as done for diagnostic studies.
Methods :
We performed an observational retrospective study from 2020 to 2015. We had local Ethics Committee approval and followed the tenets of the Helsinki declaration. Patients with retinographies, OCT, OCTA, visual fields, intraocular pressure (IOP) measurement, and complete ophthalmology records were included. A glaucoma expert reviewed the medical records to classify patients with glaucoma and those without glaucoma. Patients with systemic or other ophthalmologic pathologies and glaucoma-suspects were excluded, as well as incomplete records. Low-quality OCTA images were excluded (with strength signal index ≤50, media opacities, or fixation artifacts). We built a deep-learning software using the OCTA scans, using 90% of them to train its neural network to distinguish the images between glaucoma and non-glaucoma. The remaining ten percent were used for validation. We used the TensorFlow (v.1.15.2) encoder library, followed by several layers of mathematical operations and the RMSProp optimizer, and a learning rate of 0.001.
Results :
We included 262 patients, from which 40 were healthy controls, and 222 were glaucomatous patients. The AI system successfully discriminated glaucoma from healthy eyes from OCTA scans, with a sensitivity of 99.55%, a specificity of 92.5%, and an AUROC of 85%.
Conclusions :
Despite a small database, the AI discriminating results were promising in this pilot study. Associating OCTA scans to AI may be an effective strategy to help ophthalmologists diagnose glaucoma. More extensive multi-centric studies are needed to improve deep learning algorithms and detect glaucoma better.
This is a 2021 ARVO Annual Meeting abstract.