Title : Computer assisted quantifiable diagnostic tool for diabetic retinopathy
Abstract:
Rising prevalence of diabetic retinopathy (DR) and diabetic macular edema represents significant burden in the society and raises challenges both in establishing effective screening and in enlargening the screening scope and mitigating the risk of DR progression. Our goal is to provide the professional community with a tool which should address these present needs. The aim of the talk is introduction of the of the unique computer-assisted diabetic retinopathy (DR) screening and diagnostic tool, which is based on novelty methods of supervised learning, convolutional neural network, and deep machine learning (DML). The proprietary design of the decision algorithm, amended by standard DML protocols will be demonstrated.
The development and design of this specialized software was started in 2018 with many cooperating partners and use of supercomputing powers. The diagnostic tool was trained on thousands of retina images with multiple defined retinal pathologies. The anonymized images were firstly described and annotated by retinal specialists in the reading centers with a proprietary retina-optimized marker tool. Over a dozen different characteristic DR pathologies were selected and over > 200.000 annotations were individually marked by the retina specialists. The images were double-checked and cross annotated and ground truth was established. Then the images were delivered to the deep machine learning programmes. Almost 20 different neural networks were created and a unique algorithm for aggregation of all artificial intelligence outputs was developed.
Our diagnostic and screening software tool provides a very high accuracy rate in recognizing individual pathologies, classification of DR and mistake self-detection to dramatically minimize false positive/negative results.
The potential outcomes of our tool are twofold – firstly - the ability to screen a wide population in risk of diabetic retinopathy, secondly and most importantly to quantify pathologies on each image of the retina. This will allow disease progression tracking in the first phase, and once enriched by meta data it will eventually lead to quantifiable decision making of any medical professionals.
Cooperating partners of the project are Government of the Czech Republic (grant funding), University Eye Clinics (Charles University, Ostrava University) and a network of local reading centers with over 20 cooperating retina specialists. From a computing power perspective we used the supercomputing center provided by Technical University Ostrava, IT4 Innovation Department (grant funding by European Union).
Audience Take Away:
- New insight into the area of AI applied to digital retina image analysis (quantification, repeatability, tracking).
- What the state of the art generation of the algorithm may provide to the professional community (speed of analysis, user friendly interface, capacity limit removed, cost & time efficiency, accuracy, interpretability).
- Quantification, time tracking, and predictive function of optional interventions (pharma, laser, surgery).
