TCU-Net: Transformer Embedded in Convolutional U-Shaped Network for Retinal Vessel Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography angiography (OCTA) is crucial for visualizing retinal vasculature and diagnosing eye diseases.
• Limitations in current convolutional neural networks hinder precise microvascular segmentation in OCTA images.
• Accurate segmentation is vital for early detection and management of ophthalmic conditions.

Purpose of the Study:

• To introduce TCU-Net, an end-to-end transformer-based network for improved OCTA retinal vessel segmentation.
• To address the loss of vascular features in convolutional operations by integrating a cross-fusion transformer module.
• To enhance the fusion of multiscale features and fine-grained details using a channel-wise cross-attention mechanism.

Main Methods:

• Developed TCU-Net, a novel network architecture incorporating a cross-fusion transformer module and a channel-wise cross-attention module.
• Replaced U-Net's skip connections with a transformer module to enrich multiscale vascular features.
• Evaluated TCU-Net on the Retinal OCTA Segmentation (ROSE) dataset (ROSE-1 and ROSE-2).

Main Results:

• TCU-Net achieved high accuracy and AUC values on the ROSE-1 dataset (e.g., 0.9230 accuracy for SVC, 0.9912 for DVC).
• On the ROSE-2 dataset, TCU-Net demonstrated strong performance with 0.9454 accuracy and 0.8623 AUC.
• Experimental results indicate superior vessel segmentation performance and robustness compared to state-of-the-art methods.

Conclusions:

• TCU-Net effectively segments retinal vessels in OCTA images, outperforming existing approaches.
• The proposed transformer-based architecture enhances the extraction and fusion of microvascular information.
• TCU-Net shows significant potential for improving the diagnosis and management of ophthalmic diseases through advanced image analysis.