TransDualSegNet: transformer dual segment network for retinal vasculature segmentation in OCT.

Background: The early diagnosis and treatment of eye diseases is crucial for preserving vision, and optical coherence tomography (OCT) has become a common non-invasive technique for eye examinations. In particular, retinal vessel segmentation in OCT images is a critical task that provides detailed information such as vessel diameter and wall thickness, serving as important diagnostic aids for ophthalmologists. This study aimed to develop an efficient and accurate retinal vessel segmentation method in OCT images to improve the diagnosis and treatment of eye diseases. Methods: Transformer Dual Segment Network (TransDualSegNet) involves several key steps: firstly, the OCT images are preprocessed through denoising and brightness adjustment to enhance the performance of subsequent segmentation algorithms. Secondly, we introduce an encoder-decoder model with the Pyramid Vision Transformer as the encoding layer, which automatically learns the feature representation of retinal vessels to improve segmentation accuracy and robustness. The decoder utilizes a dual-branch convolutional network that incorporates and emphasizes the edge information of retinal vessel OCT images as complementary information for vessel segmentation. Finally, a feature fusion module (FFM) is applied to fuse the global features and edge features, which are fed into the edge extraction module to generate the final edge probability map and obtain the segmentation results. Results: Experimental results on the OCT vessel segmentation dataset showed that the proposed model outperformed U-Net, Attention U-Net, and Trans U-Net, with accuracy (ACC) 91.13%±0.11%, sensitivity (SE) 90.02%±0.38%, Jaccard similarity (JS) 93.92%±0.13%, and Dice coefficient (DICE) 90.26%±0.25% all higher than comparisons. The model has 9.82 M parameters, falling between the comparison models, and a forward inference time of 20.3 ms, faster than Attention U-Net and Trans U-Net, balancing high accuracy and efficiency. Ablation experiments showed that it outperformed all sub-models in metrics such as ACC, SE, specificity (SP), Intersection over Union (IOU), and area under the curve (AUC), with IOU reaching 93.92%±0.13% and AUC 93.04%±0.34%, verifying the synergistic effectiveness of each component. Conclusions: TransDualSegNet, with its innovative architecture and outstanding segmentation performance, offers a novel solution for retinal vessel segmentation. Although there is room for optimization, its technical advantages and clinical potential provide valuable insights for advancing medical image processing and ophthalmic diagnostic technologies. Precise retinal vessel segmentation aids in diagnosing diabetic retinopathy and hypertensive retinopathy, facilitating automatic measurement of parameters such as vessel diameter and wall thickness to offer more accurate treatment recommendations for patients.