DETECTION OF KERATOCONUS DISEASE DEPENDING ON CORNEAL TOPOGRAPHY USING DEEP LEARNING

Abstract

Keratoconus is a disease that ML has contributed much in its diagnosis and management. It is not a widely prevalent disease, with a research gap caused by the absence of standardized datasets for model training and evaluation.  This work presents a novel dataset, which strengthens the CNN model's resilience and creates standards for assessing keratoconus diagnostic techniques. The research depends on data of patients examined at Jenna Ophthalmic Center in Baghdad. The proposed system works on three stages: pre-processing, feature extraction, and classification with machine learning algorithms including NB, KNN, ADA, DT, and CNN deep learning. The pre-processing stage involves cropping images to retain the relevant maps, which were subjected to contrast enhancement to improve image quality. The pre-processed data is then fed into Machine Learning(ML) algorithms and Convolutional Neural Network(CNN) models, by which the four corneal maps were analyzed. The precision of the ML method was quantified, yielding a precision score of 0.79 for the AdaBoost algorithm and an impressive score of 0.99 for the suggested CNN model exemplifying its high accuracy and ability to surpass all machine learning approaches. Applying PCA for feature extraction before utilizing tradition ML algorithms and CNN helps in achieving high-accuracy results.