嵌入CBAM的改进Faster RCNN眼底微动脉瘤检测方法  

Microaneurysm detection in fundus using improved Faster RCNN with CBAM integration

作　　者：杨丽 邵虹[1] 崔文成[1] YANG Li;SHAO Hong;CUI Wencheng(School of Infomation Science&Engineering,Shenyang University of Technology,Liaoning 110870,China)

机构地区：[1]沈阳工业大学信息科学与工程学院,辽宁沈阳110870

出　　处：《长江信息通信》2024年第5期40-43,共4页Changjiang Information & Communications

摘　　要：眼底微动脉瘤检测可以有效地预防和控制糖尿病性视网膜病变,在临床应用中具有重要的意义,但该病灶的目标区域较小且存在眼底出血和其他结构的干扰,同时眼底图像存在亮度、对比度不均的问题,给检测任务带来了巨大挑战。针对此问题提出一种基于Faster RCNN网络的微动脉瘤小目标检测方法,先对数据集进行以病灶为中心的分块处理,提升目标区域的占比;再将主干网络替换为特征表达能力强的ResNet网络,并引入注意力机制,结合加入融合因子的特征金字塔模块进行多尺度特征融合,改善主干网络提取小目标特征信息的能力,增加其对目标区域的关注程度。实验结果表明,算法在E-Ophtha MA数据集上取得了良好的检测效果,精确率为91.3%,召回率为80.2%,较原模型精确率提高了13.1%,召回率提高了8%,且与其他方法相比检测效果更好。The detection of microaneurysms in fundus can effectively prevent and control diabet-ic retinopathy,and it has important clinical significance.However,the targct area of this lesion is small and there is interference from other structures such as retinal bleeding.In addition,there are problems of uneven brightness and contrast in fundus images,which bring great challenges to the detection task.To address this problem,a method of microaneurysm detection based on Faster RCNN is proposed.First,the dataset is segmented based on the lesion to improve the pro-portion of the target area.Then,the backbone network is replaced with ResNet which has strong feature expression ability,and an attention mechanism is introduced to combine with FPN that joined fusion factor to perform multi-scale feature fusion,thereby improving the ability of the backbone network to extract feature information of small targets and increasing its attention to the target area.Experimental results show that the algorithm achieves detection results on the E-OphthaMA dataset,with a precision of 91.3%,a recall rate of 80.2%,which is 13.1%higher and 8%higher than the original model in terms of precision and recall rates respectively,and better detection results compared to other methods.

关 键 词：小目标检测 Faster RCNN 微动脉瘤 注意力机制 多尺度特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]