Machine vs Deep Learning Algorithm Development for Cataract Detection Image Recognition

Abstract

Cataracts involving the eye's lens becoming cloudy are among the primary reasons for blindness in Indonesia and globally. The estimated annual cataract incidence is 0.1%, meaning one new cataract patient emerges every year among 1000 individuals. Indonesians tend to develop cataracts 15 years earlier than individuals in subtropical regions. In line with WHO recommendations in the IPCEC guidelines, empowering communities through promotive, preventive, curative, and rehabilitative efforts is crucial. This research aims to develop integrated AI-based cataract detection using GLCM (Gray Level Co-Occurrence Matrix) extraction methods and three machine learning algorithms for image recognition, which are KNN (K-Nearest Neighbour), SVM (Support Vector Machine), and CNN (Convolutional Neural Network). A non-implemented study design was employed to develop an AI system for cataract detection, utilizing 1,159 eye photos captured with smartphones and slit lamps. CNN achieved higher accuracy (95.31%) than SVM (81.39% or KNN (85.34%), as well as higher sensitivity (96.15%) than SVM (84%) or KNN (94%). The machine learning algorithm that produces the best results for cataract detection is CNN, with a performance score of specificity (95%), PPV (83%), and NPV (99%). We can utilize this cataract screening detection method to identify more cataract cases, thereby boosting the number of cataract surgical procedures.