PERFORMANCE EVALUATION OF DENSENET121 AND RESNET50 WITH CLAHE-BASED PREPROCESSING FOR DIABETIC RETINOPATHY DETECTION

Abstract

Background: Diabetic retinopathy (DR) had been identified as one of the main blindness-causing eye disorders and its early diagnosis was the key to avoiding blindness.

Objective: The goal of the study was to determine whether the application of five preprocessing techniques CLAHE + Normalization, Vessel Masking + CLAHE, Vessel Cropping + CLAHE, Edge Sharpening + CLAHE, Gamma Correction + CLAHE can be used to improve the performance of deep learning models of DR.

Methods: A reusable collection of retinal images was cleaned via the five enhancement techniques. The classification was executed through two robust CNN models, i.e. DenseNet121 and ResNet50. The performance of the models was checked in terms of accuracy, precision, recall, F1- score, and confusion matrices.

Results: The experiments showed that preprocessing proved to be highly influenced by the accuracy of the model. The combination of CLAHE + Vessel Masking on the dataset preprocessed using DenseNet121 has shown the best performance of all tested combinations, attaining an overall accuracy of 98.9% and being superior to other preprocessing algorithms.

Conclusion: The results revealed that the performance of CNN models in the process of detecting diabetic retinopathy was highly enhanced by implementing apt preprocessing. These findings pointed out the need to incorporate enhancement strategies into medical imaging workflows to develop feasible automatic screening and assist clinical decisions