Text

IMPLEMENTASI ARSITEKTUR UNET-TRANSFORMER DALAM SEGMENTASI SEMANTIK CITRA PEMBULUH DARAH ARTERI DAN VENA RETINA

---

Penilaian

---

Retinal blood vessels are crucial components of the eye's circulatory system, functioning to supply oxygen and nutrients while removing waste from retinal tissues. Retinal blood vessels are divided into two types: arteries and veins (A/V). Arteries and veins are often located close to each other, necessitating clear separation to assist medical professionals in identifying specific diseases and preventing errors in analyzing related conditions. The separation of arteries and veins can be performed using image segmentation-based technology. This study aims to conduct semantic segmentation of retinal blood vessels by combining the U-Net architecture, Vision Transformer (ViT), and Attention Gate. The proposed model employs ViT as an encoder to capture global spatial relationships, while U-Net acts as a decoder to restore image spatial details. An Attention Gate is integrated to filter relevant information from generated features. Segmentation performance was evaluated across five label classes: Background, Artery, Crossings, Vein, and Uncertain, using metrics including accuracy, sensitivity, specificity, F1-Score, and IoU. Evaluation results indicate that the proposed model achieved an average accuracy of 99.13%, demonstrating its ability to classify pixels with high alignment to ground truth. A sensitivity of 80.78% is classified as good, reflecting adequate balance in detecting True Positives (TP). Specificity of 91.69% indicates excellent performance in identifying Background pixels or True Negatives (TN). An F1-Score of 78.64% shows the model's reasonable balance in performance across classes. An average IoU of 77.34% suggests the model has not yet reached optimal performance in predicting overlapping areas with ground truth for artery and vein labels. This study demonstrates that the combination of U-Net, Vision Transformer, and Attention Gate effectively enhances the performance of semantic segmentation for retinal blood vessels, though improvements are still needed for specific labels

---

Availability

Inventory Code	Barcode	Call Number	Location	Status
2507001953	T169750	T1697502025	Central Library (Reference)	Available but not for loan - Not for Loan

Detail Information

Series Title

-

Call Number

T1697502025

Publisher

: Prodi Ilmu Matematika, Fakultas Matematika Dan Ilmu Pengetahuan Alam Universitas Sriwijaya., 2025

Collation

xvii, 65 hlm.: Ilus., tab.; 29 cm

Language

Indonesia

ISBN/ISSN

-

Classification

510.07

Content Type

-

Media Type

-

Carrier Type

-

Edition

-

Subject(s)

Matematika

Specific Detail Info

-

Statement of Responsibility

EM

Other version/related

File Attachment

Comments

---

You must be logged in to post a comment