Real Time Glaucoma Prediction Using Y-UNet Classifier via Hardware Software Co-Design SMART System

Abstract

Objectives: To present a real-time algorithm that combines Yolov5 and UNetbased CNN predictions to classify small-sized images, particularly medical images. Methods: The proposed model combines the various phases of preprocessing, object detection, segmentation and classification through Kalman filter, Yolov5, U-net based on CNN. The model is derived from the three different datasets to create a novel classification algorithm for medical data. The dataset contains images of 136 glaucoma patients and 187 healthy images. Findings: The proposed Y-UNet classifier framework is used to classify glaucoma images based on their IOP and CCT from real-time dataset collection.The proposed framework accuracy of 98.75% is achieved in the run time performance at 0.18 seconds per image. The accuracy is higher by 1.66% and runtime performance reduced by 0.03 seconds when compared with standard classification methods. Additionally, threshold optimizer has minimized the overall losses and provides the most accurate result. Novelty: The Y-UNet classifier model proposed here enables the real-time glaucoma prediction tasks to be carried out through the use of a hardware-based system known as the Sensitive Mirror Analyzer and Retina Tracker. To the best of our knowledge, for the first time, the proposed algorithm has been implemented to improve the accuracy and run time performance.