Comparative Analysis of An AI Based Diagnostic System For Automated Detection of Diabetic Retinopathy

This work is a new comparative study of machine learning algorithms for automated diabetic retinopathy (DR) detection from retinal images.It is designed and tested a novel AI-based system for diagnosis using our new Multi-Feature Weighted Ensemble (MFWE) framework that utilized multiple publicly available datasets such as MESSIDOR, KAGGLE EyePACS, APTOS 2019, and IDRiD.Our method involved extensive preprocessing of data, feature extraction, and the execution of five different algorithms for classification: Linear Regression, Random Forest, XGBoost, MLP Classifier, and Decision Tree Classifier.The novel preprocessing process involved image standardization, color normalization, noise reduction, contrast enhancement, and anatomical structure segmentation.The derived unique features such as morphological features, vessel measurements, optic disc features, texture features, color features, and wavelet-based features.Dimensionality reduction was done using Principal Component Analysis and Recursive Feature Elimination with Cross-Validation.Experiments were carried out on the Messidor dataset of 2,302 samples with 19 features extracted.Models were evaluated by using such metrics as accuracy, precision, recall, F1-score, and ROC curve.Our implementation of Feature-Weighted Ensemble turned out to be distinctive with outstanding performance compared to traditional methods.Treebased models yielded best results with XGBoost showing the highest F1-score (0.911) and accuracy (0.907), closely followed by Decision Tree (F1: 0.901, accuracy: 0.896).The novelty in our approach comes from combining MFWE with best-performing hyperparameters, illustrating improved precision-recall balance along with lower overfitting versus baseline implementations.The good performance of relatively naive models indicates our extracted features worked well in selecting informative patterns towards DR classification.Our results show that our novel approach integrating tree-based models with feature-weighted ensemble methods is very effective for DR detection.This research adds a new method to the construction of trustworthy AI-based screening tools that may help ophthalmologists in early detection of diabetic retinopathy.