改进型DeepLabV3+的糖尿病眼底病变分割  

Diabetic fundus lesion segmentation by improved DeepLabV3+

作　　者：马晓普[1] 刘文涛 李贺 MA Xiaopu;LIU Wentao;LI He(School of Computer Science and Technology,Nanyang Normal University,Nanyang 473061,Henan China;School of Life Sciences and Agricultural Engineering,Nanyang Normal University,Nanyang 473061,Henan China)

机构地区：[1]南阳师范学院计算机科学与技术学院,河南南阳473061 [2]南阳师范学院生命科学与农业工程学院,河南南阳473061

出　　处：《华中科技大学学报（自然科学版）》2024年第5期90-97,共8页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62002180);南阳师范学院“卧龙学者”奖励计划支持项目基金资助项目.

摘　　要：针对糖尿病视网膜眼底病变多类分割难及小病灶识别率低的问题,提出一种融合注意力机制与改进型DeepLabV3+的糖尿病视网膜眼底病变多类病症分割方法.该方法首先通过使用MobileNetV2网络提取病灶特征,减少参数量并提高算法训练速度;接着通过优化空洞空间卷积金字塔池化中的空洞卷积层数与空洞率,以提高捕获小病灶特征的能力;然后基于DeepLabV3+网络模型进行改进,借助坐标注意力机制感知病灶方向和位置信息,从而提高识别精度;最后采用FGADR和IDRiD数据集分别训练和测试所提出的模型.实验结果表明:所提出方法的平均交并比(MIoU)指标为73.75%,具有较高的分割精度,验证了模型有效性.Aiming at the problem of difficulties of multi-class segmentation and low recognition rate for small lesions in diabetic retinopathy,a method combining attention mechanism with an improved DeepLabV3+model was proposed for multi-class lesion segmentation.First,the MobileNetV2 network was employed to extract lesion features,reducing parameter count and enhancing training speed.Subsequently,the dilation convolution layers and dilation rates in the atrous spatial pyramid pooling were optimized to improve the capability of capturing features of small lesions.Then,improvements were made to the DeepLabV3+model by incorporating a coordinate attention mechanism to perceive lesion direction and position information,thereby enhancing recognition accuracy.Finally,the proposed model was trained and tested on the fine-grained annotated diabetic retinopathy(FGADR)and Indian diabetic retinopathy image dataset(IDRiD).Experimental results show that the proposed method achieves a mean intersection over union(MIoU)metric of 73.75%,showcasing high segmentation accuracy,and confirming the effectiveness of the model.

关 键 词：糖尿病视网膜眼底病变 深度学习 DeepLabV3+网络 坐标注意力 多类分割 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]