HealthLens AI: Revolutionizing Glaucoma Detection with Cutting-Edge AI

Mission

“Revolutionizing glaucoma detection with cutting-edge AI: precise, fast, and non-invasive analysis of fundus images for better patient outcomes. Committed to transforming eye care, making it more accurate and accessible.”

Value Proposition 

The prevalence of glaucoma, the second leading cause of blindness globally, underscores the urgency of early and accurate detection. By 2040 alone, the WHO estimates that approximately 112 million people will be affected by this disease. While fundus photography is instrumental in diagnosing such eye diseases, interpreting the resulting images demands specialized knowledge, often scarce in primary care settings, as well as specialized tests and measurements taken from ocular structural characteristics. 

This project aims to bridge this gap by leveraging advanced computer vision techniques to provide initial risk determinations of glaucoma as well as to identify and measure key structural characteristics non-invasively using industry standard fundus photography images. With our product, we aim to provide medical practitioners with a robust evidence package enabling more optimal decisions and ultimately diagnoses of glaucoma. 

We envision our solution to be a state of the art AI assistant for ophthalmologists and optometrists. This proof of concept establishes a baseline that can process ocular imagery for Glaucoma Diagnosis alongside associated structural characteristic detection. The ultimate goal is to extend the scope to provide a comprehensive risk panel and automated measurements across a variety of target diseases, ocular structures, and image modalities beyond our MVP.

Approach 

We are developing a novel approach that uses state of the art deep learning methods in computer vision to initially detect the likelihood of glaucoma using full color fungus photography while also identifying key structural characteristics such as the optic disc, optic cup, and ocular blood vessel network. Using initially detected risks of glaucoma alongside derived structural metrics (e.g. vertical and horizontal cup to disk ratio), we utilize ensemble methods to improve predictions of glaucoma risk above what a single model is able to achieve on its own. 

Key Tasks 

• Detecting risk of glaucoma 

• Detecting Primary Ocular Structures (Optic Disc, Optic Cup, Blood Vessel Network)

• Autonomously computing key metrics on derived ocular structures (cup to disc ratios, relative size of structures, etc) 

Dataset 

• SMDG-19 Dataset, the largest open-source collection of annotated fundus photography images with glaucoma / non-glaucoma labels. Several subsets of this dataset also contained labeled segmentation tasks for our structural identification tasks.

• Approximately 12500 images 

• Key Constraint: Patient demographic data as well as unique patient identifiers are sparse across this dataset

Experimentation

Achieved State of The Art result with a novel Segmentation Informed Classification approach using CNN and Vision Transformers, results further enhanced with a self-hosted LLM. Deployed in a scalable architecture, this product can be seamlessly integrated with any device capable of taking a color fundus image.