<![CDATA[Diabetic retinopathy classification using fundus images increasingly relies on deep learning, particularly Convolutional Neural Networks (CNNs). However, the use of high-dimensional retinal images in the processing pipeline may increase computational requirements. This study proposes a DLSTM-121-based approach for diabetic retinopathy classification by applying Principal Component Analysis (PCA) to augmented retinal images before U-Net-based segmentation. The main methodological contribution is the placement of PCA before the segmentation stage, allowing dimensionality reduction to be performed before the images are processed by the subsequent segmentation and classification models. DLSTM-121 combines DenseNet-121 and Long Short-Term Memory (LSTM) to extract spatial features and model dependencies among the extracted features. The proposed pipeline consists of data augmentation, PCA, U-Net-based retinal image segmentation, and DLSTM-121 classification. Four PCA configurations using 25, 50, 75, and 100 principal components were evaluated. The results showed that 75 principal components retained approximately 99% of the image information and achieved the best classification performance, with 96.92% accuracy, 100% sensitivity, 86.67% specificity, 98.18% F1-score, and 99% AUC. This configuration required 1,435.59 s of training time and 261.76 MB of peak GPU memory, compared with 1,456.18 s and 232.41 MB for the configuration without PCA. The inference times were 177.22 and 173.20 ms/image, respectively, while both configurations required 12.713 GFLOPs. These results demonstrate that PCA can reduce the dimensionality of the retinal image representation while maintaining high classification performance. The study used 129 retinal images from a single dataset, which limits the generalizability of the findings. Future research should evaluate the proposed approach using larger and more diverse retinal datasets.]]>
Copyrights © 2026
