Liu Zerui, Du Junliang, Dai Weisi, Zhu Wenke, Ye Ziqing, Li Lin, Liu Zewei, Hu Linan, Chen Lin, Sun Lixiang
Faculty of Electronic Information and Physics, Central South University of Forestry and Technology, Changsha, 410004, Hunan, China.
Artificial Intelligence Institute, Shanghai Jiao Tong University, Shanghai, 200240, China.
Sci Rep. 2025 Jan 6;15(1):900. doi: 10.1038/s41598-024-84901-3.
Accurate segmentation of retinal blood vessels from retinal images is crucial for detecting and diagnosing a wide range of ophthalmic diseases. Our retinal blood vessel segmentation algorithm enhances microfine vessel extraction, improves edge texture clarity, and normalizes vessel distribution. It stabilizes neural network training for complex retinal vascular features. Channel-aware self-attention (CAS) improves microfine vessel segmentation sensitivity. Heterogeneous adaptive pooling (HAP) facilitates accurate vessel edge segmentation through multi-scale feature extraction. The ghost fully convolutional Rectified Linear Unit (GFCReLU) module in the output convolutional layer captures deep semantic information for better vessel localization. Optimization training with Sparrow-Integrated Lion Optimization Algorithm (SLOA) employs sparrow stochastic updating and annealing to fine-tune parameters. The results of the experiments on our homemade dataset and three public datasets are as follows: Mean Intersection over Union (MIoU) of 80.61%, 76.14%, 76.90%, 74.11%; Dice coefficients of 78.97%, 72.51%, 72.84%, 68.93%; and accuracies of 94.83%, 95.74%, 96.67%, 95.81% respectively. The model effectively segments retinal blood vessels, offering potential for diagnosing ophthalmic diseases. Our dataset is available at https://github.com/ZhouGuoXiong/Retinal-blood-vessels-for-segmentation .
从视网膜图像中准确分割出视网膜血管对于检测和诊断多种眼科疾病至关重要。我们的视网膜血管分割算法增强了微血管提取能力，提高了边缘纹理清晰度，并使血管分布正常化。它稳定了针对复杂视网膜血管特征的神经网络训练。通道感知自注意力（CAS）提高了微血管分割的灵敏度。异构自适应池化（HAP）通过多尺度特征提取促进了准确的血管边缘分割。输出卷积层中的幽灵全卷积整流线性单元（GFCReLU）模块捕获深度语义信息以实现更好的血管定位。使用麻雀集成狮子优化算法（SLOA）进行优化训练采用麻雀随机更新和退火来微调参数。在我们自制的数据集和三个公共数据集上的实验结果如下：平均交并比（MIoU）分别为80.61%、76.14%、76.90%、74.11%；骰子系数分别为78.97%、72.51%、72.84%、68.93%；准确率分别为94.83%、95.74%、96.67%、95.81%。该模型有效地分割了视网膜血管，为眼科疾病的诊断提供了潜力。我们的数据集可在https://github.com/ZhouGuoXiong/Retinal-blood-vessels-for-segmentation获取。
