Skripsi
MODIFIKASI SWIN-UNET DENGAN COORDINATE ATTENTION, DUAL ATTENTION BLOCK, DAN SPATIAL ATTENTION MODULE UNTUK SEGMENTASI SEMANTIK PEMBULUH DARAH RETINA
Abnormalities in retinal blood vessels can be used as indicators for detecting various eye diseases, therefore retinal blood vessel image segmentation is required. Manual segmentation is time-consuming, thus an automatic segmentation method based on deep learning is needed to assist the segmentation process of retinal blood vessel images. This study aims to obtain the performance results of semantic segmentation of retinal blood vessels using a modified Swin-Unet architecture with the application of Coordinate Attention (CA), Dual Attention Block (DAB), and Spatial Attention Module (SAM) on LES-AV images. The proposed modifications focus on three main parts, namely CA to preserve spatial positional information in the encoder, DAB to focus attention on important features in the bottleneck, and SAM to maintain spatial details in the decoder. Retinal images are processed through a preprocessing stage using Contrast Limited Adaptive Histogram Equalization (CLAHE) and data augmentation in the form of flipping, rotation, and zooming to improve data quality and diversity. The model is trained to produce three segmentation classes, namely artery, vein, and background. The results of this study show that the modified Swin-Unet architecture with CA, DAB, and SAM provides better segmentation performance compared to other standard models. Model performance evaluation yields an accuracy of 99.32%, sensitivity of 94.91%, specificity of 97.52%, precision of 95.07%, F1-score of 94.99%, and Intersection over Union (IoU) of 90.67%. These results indicate that the modified Swin-Unet architecture with CA, DAB, and SAM is capable of improving the performance of retinal blood vessel segmentation.
No other version available