摘要： 在临床诊断中，眼底硬性渗出物的检测结果是判断糖尿病视网膜病变程度的重要参考。现有眼底硬性渗出物检测模型通过加深网络层数以有效分割硬性渗出物的病灶特征，但是容易产生冗余卷积单元且难以准确提取全部有效特征，影响整体分割性能。提出一种融合交错组卷积与双重注意力机制的眼底硬性渗出物自动分割模型。利用改进的交错组卷积模块代替原始U型网络编码部分，在减少分割模型参数的同时提取更丰富的病灶特征。同时通过位置注意力模块联系局部上下文信息，捕获更广泛的感受野以及更深层次的病灶特征，利用通道注意力模块增加提取关键特征的通道权重，提升重要特征的可辨别性。实验结果表明，该模型在e-Ophtha EX数据集上灵敏度、精确度和F-Score分别为91.43%、86.49%和87.32%，在DIARETDB1数据集上灵敏度、特异性和准确性分别达到97.83%、96.16%和97.51%，能够有效改善原始U型网络对眼底硬性渗出物的分割效果。

关键词: 糖尿病视网膜病变, 硬性渗出物, 交错组卷积, 位置注意力机制, 通道注意力机制, U型网络

Abstract: In clinical diagnosis, the detection results of the fundus Hard Exudate(HE) are an important reference to assess the degree of Diabetic Retinopathy(DR).Although existing fundus HE detection models can effectively segment the features of HE lesions by deepening the number of network layers, redundant convolution units are easily generated.This significant shortcoming in current fundus HE detection models degrades the overall segmentation performance.It is, therefore, difficult to accurately extract all the significant features.Thus, an automatic fundus HE segmentation model based on Interleaved Group Convolution (IGC) and fusion of dual attention mechanisms are proposed. The improved IGC module replaces the original U-shaped network coding component and extracts richer lesion features while reducing the segmentation model parameters.The position attention module contacts the local context information to capture a broader receptive field and deeper lesion features.The channel attention module increases the channel weighting in extracting key features and improves the discriminability of important features.The experimental results show that the model's senstivity, positive predictive value, and F-Score on the e-Ophtha dataset are 91.43%, 86.49% and 87.32% respectively.Additionally, the senstivity, specificity, and accuracy on the DIARETDB1 dataset are 97.83%, 96.16% and 97.51% respectively.This effectively enhances the segmentation effect of the original U-shaped network on the fundus HE.

Key words: Diabetic Retinopathy(DR), Hard Exudate(HE), Interleaved Group Convolution(IGC), position attention mechanism, channel attention mechanism, U-shaped network

中图分类号: 

• TP391