Analyzing the effect of preprocessing methods on deep learning-based classification of diabetic retinopathy degrees

Diabetic retinopathy (DR) is a severe complication of diabetes and an ocular disease that can lead to vision loss if left untreated. Early diagnosis and accurate grading are critically important for effective treatment processes. In this context, deep learning-based classification methods offer promising approaches for DR detection. However, the success of deep learning models heavily depends on the preprocessing techniques applied. This thesis investigates the effects of various preprocessing methods on the deep learning-based classification of different DR degrees. In the scope of this thesis study, the effects of various preprocessing techniques (Gaussian blur, thresholding, contour detection, CLAHE, and median blur) on classification performance are analyzed. The preprocessing methods are applied in two distinct steps. In the first preprocessing step, Gaussian blur, thresholding, and contour detection methods are applied to the dataset and trained. Subsequently, the images from the first preprocessing step undergo the second preprocessing step, where CLAHE and median blur methods are applied and then trained. Transfer learning-based models such as EfficientNetB5, VGG16, and VGG19 are utilized, and the results obtained after each preprocessing step are compared. Experiments are conducted on two different datasets, APTOS and Diabetic Retinopathy (resized), to evaluate the impact of preprocessing methods on performance metrics such as classification accuracy, precision, recall, and F1 score. For the APTOS dataset, the accuracy rates of transfer learning methods before any preprocessing were observed as EfficientNetB5 (96.09%), VGG16 (95.18%), and VGG19 (95.25%). After applying the first preprocessing step (Gaussian blur – thresholding – contour detection), the accuracy rates were EfficientNetB5 (95.49%), VGG16 (95.40%), and VGG19 (95.05%). Following the second preprocessing step (CLAHE – Median blur), the accuracy rates were EfficientNetB5 (93.63%), VGG16 (95.18%), and VGG19 (94.89%). For the Diabetic Retinopathy (resized) dataset, the accuracy rates of transfer learning methods before any preprocessing were EfficientNetB5 (87.95%), VGG16 (89.72%), and VGG19 (89.72%). After the first preprocessing step (Gaussian blur – thresholding – contour detection), the accuracy rates were EfficientNetB5 (87.77%), VGG16 (89.54%), and VGG19 (89.54%). Following the second preprocessing step (CLAHE – Median blur), the accuracy rates were EfficientNetB5 (87.21%), VGG16 (89.54%), and VGG19 (89.54%). Additionally, the role of preprocessing methods in the models' accuracy and loss curves is examined. The results indicate that the preprocessing steps provide partial improvements in model performance but do not significantly enhance classification accuracy. These findings suggest that preprocessing methods in deep learning-based classification processes may not always yield substantial performance improvements. However, they can be crucial for improving image quality and contributing to model stability. This study offers a comprehensive framework for identifying suitable preprocessing strategies for deep learning models used in DR diagnosis. The thesis aims to contribute to the development of deep learning-based solutions that enable faster and more reliable DR diagnosis.