| 摘 要: 为提高脉络膜分割算法对多尺度特征的提取和特征融合问题,提出了一种基于多尺度特征提取和自适应特征融合的OCT图像脉络膜分割方法。首先,通过多分支结构和异构卷积策略构建特征提取模块,提高网络对局部特征和全局特征的提取能力;其次,采用空间注意力机制重构跳跃连接,将不同的权重应用于不同级别的特征,强化对提取特征的区分能力。实验结果表明,所提方法在分割精度、灵敏度、MIoU和Dice系数分别达到了99.6%、96.8%、94.7%、94.8%,相较于U-Net提高了0.5个百分点、3.4个百分点、1.6个百分点、1.9个百分点,验证了该方法在提升脉络膜分割任务性能的有效性。 |
| 关键词: 脉络膜 深度学习 特征提取 感受野 U-Net 损失函数 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 国家自然科学基金资助项目(61663005) |
|
| Choroid Segmentation in OCT Images Based on Multi-Scale Feature Extraction and Adaptive Feature Fusion |
|
CHEN Dejie1, CHEN Xiangping1, GU Hao2, ZHAO Su2, JIANG Hao2
|
(1. College of Electrical Engineering, Guizhou University, Guiyang 550025, China;
2. Department of Ophthalmology, Affiliated Hospital of Guizhou Medical University, Guiyang 550004, China)
2366984419@qq.com; xp.eechen@gzu.edu.cn; guhao@gmc.edu.cn; ximi520@sina.com; kidd5jh0@sina.com
|
| Abstract: To enhance the extraction and fusion of mult-i scale features in choroid segmentation algorithms, a method for choroid segmentation in OCT images based on multi-scale feature extraction and adaptive feature fusion is proposed. First, a feature extraction module is constructed using a multi-branch structure and heterogeneous convolution strategy to improve the network’s ability to capture both local and global features. Second, a spatial attention mechanism is employed to redesign skip connections, applying different weights to features at various levels to strengthen the discriminative capability of the extracted features. Experimental results demonstrate that the proposed method achieves segmentation accuracy, sensitivity, MIoU, and Dice coefficient of 99.6% , 96.8% , 94.7% , and 94.8% ,respectively, representing improvements of 0.5 percentage points, 3.4 percentage points, 1.6 percentage points, and 1.9 percentage points compared to the U-Net. These results validate the effectiveness of the proposed method in enhancing the performance of choroid segmentation tasks. |
| Keywords: choroid deep learning feature extraction receptive field U-Net loss function |
