基于CMSA-Unet的脉络膜血管分割

Choroidal vessel segmentation based on CMSA-Unet

为了自动精准地分割脉络膜血管以辅助眼科治疗,提出一种基于脉络膜形态拉伸增广(choroid morphology stretch,CMS)结合注意力机制的U-Net框架CMSA-Unet(choroid morphology stretch ATT-Unet).模型先通过数据预处理模块和数据增广模块对原始光学相干断层扫描(optical coherence tomography,OCT)图像进行处理,增强脉络膜的形态特征,并扩充有限的数据集;再在每个编码器中设置卷积注意力模块(convolutional block attention module,CBAM),使模型关注目标分割区域,提升分割效果.消融实验表明,CMS可增强脉络膜特征,CMS结合CBAM模块能有效提高模型对脉络膜血管的分割效果,其IoU、F1分数、灵敏度分别比基线模型U-Net提高了2.9%、2.0%、5.6%.相较于同类模型,CMSA-Unet准确率更高,更适用于脉络膜分割任务.

In order to automatically and accurately segment choroidal vessels for ophthalmic treatment, this paper proposes a U-Net framework CMSA-Unet(choroid morphology stretch ATT-Unet) based on choroid morphology stretch(CMS) combined with attention mechanism. Firstly, the original optical coherence tomography(OCT) images are processed by the data preprocessing module and data stretching module to enhance the morphological characteristics of the choroid and expand the limited data set. Then a convolutional blockattention module(CBAM) is set in each encoder to enable the model focus on the target segmentation region and improve the segmentation effect. Ablation experiments show that CMS can enhance choroidal features, and CMS combined with CBAM module can effectively improve the segmentation effect of the model on choroidal vessels, and the IoU, F1 score, and sensitivity are 2.9%, 2.0%, and 5.6% higher than that of the baseline model U-net, respectively. Compared with similar models, CMSA-Unet has higher accuracy and is more suitable for choroidal segmentation tasks.