Li Wei, Cao Li, Deng He
School of Electrical and Electronic Engineering, Wuhan Polytechnic University, Wuhan, China.
School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan, China.
Front Artif Intell. 2025 Jul 17;8:1463233. doi: 10.3389/frai.2025.1463233. eCollection 2025.
Since optical coherence tomography angiography (OCTA) is non-invasive and non-contact, it is widely used in the study of retinal disease detection. As a key indicator for retinal disease detection, accurate segmentation of foveal avascular zone (FAZ) has an important impact on clinical application. Although the U-Net and its existing improvement methods have achieved good performance on FAZ segmentation, their generalization ability and segmentation accuracy can be further improved by exploring more effective improvement strategies.
We propose a novel improved method named Feature-location Attention U-Net (FLA-UNet) by introducing new designed feature-location attention blocks (FLABs) into U-Net and using a joint loss function. The FLAB consists of feature-aware blocks and location-aware blocks in parallel, and is embed into each decoder of U-Net to integrate more marginal information of FAZ and strengthen the connection between target region and boundary information. The joint loss function is composed of the cross-entropy loss (CE loss) function and the Dice coefficient loss (Dice loss) function, and by adjusting the weights of them, the performance of the network on boundary and internal segmentation can be comprehensively considered to improve its accuracy and robustness for FAZ segmentation.
The qualitative and quantitative comparative experiments on the three datasets of OCTAGON, FAZID and OCTA-500 show that, our proposed FLA-UNet achieves better segmentation quality, and is superior to other existing state-of-the-art methods in terms of the MIoU, ACC and Dice coefficient.
The proposed FLA-UNet can effectively improve the accuracy and robustness of FAZ segmentation in OCTA images by introducing feature-location attention blocks into U-Net and using a joint loss function. This has laid a solid theoretical foundation for its application in auxiliary diagnosis of fundus diseases.
由于光学相干断层扫描血管造影（OCTA）具有非侵入性和非接触性，因此在视网膜疾病检测研究中被广泛应用。作为视网膜疾病检测的关键指标，黄斑无血管区（FAZ）的准确分割对临床应用具有重要影响。尽管U-Net及其现有的改进方法在FAZ分割上取得了良好的性能，但通过探索更有效的改进策略，其泛化能力和分割精度仍可进一步提高。
我们提出了一种名为特征定位注意力U-Net（FLA-UNet）的新型改进方法，通过将新设计的特征定位注意力块（FLAB）引入U-Net并使用联合损失函数。FLAB由并行的特征感知块和位置感知块组成，并嵌入到U-Net的每个解码器中，以整合更多FAZ的边缘信息，并加强目标区域与边界信息之间的联系。联合损失函数由交叉熵损失（CE损失）函数和骰子系数损失（Dice损失）函数组成，通过调整它们的权重，可以综合考虑网络在边界和内部分割上的性能，以提高其对FAZ分割的准确性和鲁棒性。
在OCTAGON、FAZID和OCTA-500这三个数据集上进行的定性和定量对比实验表明，我们提出的FLA-UNet实现了更好的分割质量，并且在平均交并比（MIoU）、准确率（ACC）和骰子系数方面优于其他现有的先进方法。
所提出的FLA-UNet通过将特征定位注意力块引入U-Net并使用联合损失函数，可以有效提高OCTA图像中FAZ分割的准确性和鲁棒性。这为其在眼底疾病辅助诊断中的应用奠定了坚实的理论基础。
