Automated Detection of Retinopathy of Prematurity Using Quantum Machine Learning and Deep Learning Techniques

Abstract

Retinopathy of prematurity (ROP) is a vasoproliferative retinal disease that affects premature infants and causes permanent blindness if left untreated. Automated retinal diagnosis from the Retinal fundus images aid in the early detection of many pathological conditions. The low-level statistical features used in literatures have not provided the complete ROP-specific profile, and hence it has to be replaced by high-level features. The proposed system involves extracting Scale Invariant Feature Transform (SIFT) - Speeded Up Robust Features (SURF) combined high-level features from the SegNet segmented retinal vessels and classified using the Quantum Support Vector Machine (QSVM) classifier. This study aims (i) to segment retinal vessels from the acquired fundus images using SegNet and extract their features using the SURF and SIFT Feature Extraction method, (ii) to classify the Normal and ROP retinal vessels using four classical machine learning classifiers such as Support Vector Machine (SVM), Reduced Error Pruning (REP) tree, K-Star, and LogitBoost and Quantum SVM classifier, (iii) to develop a novel transformer-based Swin-T ROP model to classify ROP from normal Neonatal fundus images, (iv) to compare the performance characteristics of the proposed QSVM model with the Resnet50, DarkNet19, and classical machine learning classifiers. The study is conducted using 200 fundus images, including 100 normal and 100 ROP-positive neonatal retinal images. The machine learning classifiers such as SVM, REP Tree, K-Star, and Logit Boost Classifiers attained accuracy of 86.7%, 75%, 74%, and 76.5%, respectively, in classifying ROP from normal retinal images. The deep learning networks such as ResNet50 and DarkNet19 classified ROP from normal fundus images with an accuracy of 92.87% and 89%, respectively. The Quantum machine learning classifier outperforms the classical machine learning classifiers, Pre-trained Convolutional Neural Networks (CNN) and SwinT-ROP in terms of classification accuracy (95.5%), sensitivity (93%), and specificity (98%). The proposed system accurately diagnoses ROP from the neonatal fundus images and could be used in point-of-care diagnosis to access diagnostic expertise in underserved regions.

Keywords

• QSVM
• transformer
• transfer learning
• quantum classifier
• retinal image processing