摘要
视网膜血管的自动分割技术有助于早期诊断和治疗与视网膜相关的疾病。由于视网膜血管结构复杂且精细,眼底图像存在着低对比度、光照不均以及病理性渗出物等因素的干扰,导致该任务仍然具有挑战性。针对该任务主流框架U-Net中未考虑全局语义依赖关系以及编码器和解码器之间的语义鸿沟问题,提出了一种同尺度和跨尺度增强的U-Net模型。从两个角度对该模型进行设计:对于同一尺度的编码-解码层,一种空间增强的自注意力机制被嵌入到每个编码层中以增强模型的全局空间聚合能力,并进一步将其拓展到解码端来缓解解码过程中上采样操作带来的信息丢失等问题;对于不同尺度的编码-解码层,引入了一种新颖的跨尺度融合模块,通过动态地选择最深层中丰富的特征信息来增强与其它层之间的语义交互,从而进一步弥合编码器和解码器之间的语义鸿沟。在DRIVE、CHASE_DB1和STARE三个视网膜标准数据集上进行了实验验证,实验结果表明I2A-Net能有效地分割出视网膜血管结构,相比与基线模型,在各项评价指标上均取得了较高的提升。
Automatic segmentation of retinal vessels facilitates the diagnosis and treatment of retina-related diseases.The task is challenging because of the complex structure of retinal vessels and interferences in fundus images,such as low contrast,uneven illumination,and pathological exudates.To address the lack of global semantic dependency modeling and the semantic gap between encoders and decoders in the U-Net for this task,an Intra-and inter-scale augmented U-Net(I2A-Net) is proposed.I2A-Net is designed based on two perspectives:for the intra-scale encoding-decoding layer,a Spatial Enhanced Self-attention mechanism is integrated in each coding layer to enhance the global spatial aggregation capability,and further developed into the decoder to alleviate the information loss caused by up-sampling operations;For inter-scale encoding-decoding layer,a novel Cross-scale Fusion module is introduced to boost the semantic interaction with other layer by dynamically leveraging the rich feature information of the deepest layer,thus further alleviating the semantic gap between encoder and decoder.The experiments on three retinal standard datasets,DRIVE,CHASE_DB1 and STARE,demonstrate that I2A-Net can effectively segment the retinal vessel structure,and compared with Baseline,I2A-Net can yield superior performance in all evaluation metrics.
作者
杨颖
岳圣斌
楚博文
全海燕
YANG Ying;YUE Shengbin;CHU Bowen;QUAN Haiyan(College of Information Engineering and Automation,Kunming University of Science and Technology;Yunnan Provincial Key Laboratory of Artificial Intelligence, Kunming University of Science and Technology)
出处
《光学技术》
CAS
CSCD
北大核心
2023年第4期487-496,共10页
Optical Technique
基金
国家自然科学基金(61861023,41364002)。
