| Popis: |
Retinal vessel segmentation is crucial for the diagnosis and monitoring of ophthalmic illnesses. Deep learning algorithms have been extensively utilized in automated segmentation to improve effectiveness and efficiency. In this paper, we introduce the use of DeepLabV3+ architecture to segment retinal blood vessels and enhance its performance by applying six different deep neural network backbones: ResNet50, DenseNet121, MobileNetV2, Xception, Xception with lower features (XceptionLF), and Xception lower features with overlapping regions (XceptionLFOR) patches. We also demonstrate the robustness of placing the Swin Transformer into the DeepLabV3+ model. The integration of XceptionLF and XceptionLFOR into the pipeline enhances the semantic segmentation of retinal images by enabling the merging of global and patch-specific features along with features from both lower and higher resolutions. The enhancements enable our proposed best model, XceptionLFOR, to obtain an 89.23% dice score, which represents a significant advancement in applying advanced deep-learning techniques for medical imaging. The XceptionLFOR model achieves a higher performance and better $F1$ score (0.49%) over the state-of-the-art for the FIVES benchmark evaluation despite using lower image resolution (256 resolution patches from 512-resolution images). The use of lower resolution balances computational efficiency with enhanced performance, enabling faster processing and deployment in resource-constrained environments. The findings in this paper point in the right direction in improving semantic segmentation for retinal vessel images, and they highlight the potential to improve early diagnosis and treatment outcomes for ocular illnesses. |
