<![CDATA[Proliferative Diabetic Retinopathy (PDR) is the most severe stage of Diabetic Retinopathy (DR), carrying the highest risk of complications. Automatic detection can help provide earlier and more accurate PDR diagnosis, but prediction accuracy may decline due to limitations in retinal images. Therefore, image enhancement techniques are often applied to improve DR classification. This study aims to detect PDR from retinal images using Convolutional Neural Networks (CNNs) and to evaluate the impact of three enhancement methods. This research method is based on a CNN architecture, including ResNet34, InceptionV2, and DenseNet121, as well as enhancement methods such as CLAHE, Homomorphic Filtering (HF), and Multiscale Contrast Enhancement (MCE). The results of this research show that CNN performance varies across architectures and enhancement methods. The highest performance was achieved using ResNet34 with HF, yielding an accuracy of 0.976, precision of 0.934, and recall of 0.904. CLAHE generally improved performance across architectures, achieving the best average accuracy of 0.953, whereas MCE decreased classification accuracy. Overall, the findings highlight the importance of selecting appropriate enhancement methods to improve PDR detection accuracy. Implementing such systems in clinical screening could help reduce the risk of vision impairment among diabetic patients.]]>
Copyrights © 2025
