A Novel Network for Choroidal Segmentation Based on Enhanced Boundary Information

Abstract

Due to the inherent imaging characteristics of Optical Coherence Tomography (OCT), the contrast between the choroid and sclera is low, presenting significant challenges in choroidal segmentation, such as subchoroidal boundary blurring and difficulty in accurate boundary delineation. To address these issues, this paper proposes an automatic choroid segmentation network, termed Boundary Enhancement Net (BENet), which enhances boundary information to facilitate precise recognition and achieve end-to-end automatic segmentation. BENet is constructed on the TransUnet backbone and integrates a boundary enhancement module, leveraging a multi-branch architecture and dilated convolutions to capture features across multiple scales. The attention mechanism is incorporated to dynamically highlight critical features, thus improving the model’s capacity to represent boundary details. Furthermore, the channel enhancement mechanism further refines the expression of salient channel futures, while the adaptive activation function improves the sensitivity of the network to boundary futures. A custom loss function, specifically designed to replace the conventional Mean Squared Error (MSE) loss, is utilized to optimize network parameters, leading to further improvements in segmentation performance. The BENet architecture achieved a Dice coefficient of 95.24%, a Hausdorff distance of 2.68, and an accuracy of 99.12%. Compared to the original TransUnet, BENet demonstrated an increase of nearly 5% in the Dice coefficient and 0.85% in accuracy. Ablation studies and results from multiple medical image datasets validate that BENet consistently delivers precise segmentation outcomes. The boundary enhancement module effectively improves the accuracy of choroidal segmentation, while the custom loss function optimizes network performance, increasing both stability and generalization capabilities. Moreover, the modular design of the boundary enhancement module ensures its portability across different segmentation tasks, making it a versatile component for integration into existing frameworks.

Keywords

• Choroid segmentation
• TransUnet
• boundary enhancement
• adaptive activation
• deep learning