基于非对称式跳跃连接的轻量化糖网病变分割算法

A Lightweight Sugar Web Lesion Segmentation Algorithm Based on Asymmetric Skip Connections

糖尿病患者视网膜病变(DR)是致盲的主要原因。眼底病变是DR的临床表现,因此,对其进行准确的识别对于疾病的早期筛查、分级和监测至关重要。本文提出了一种轻量化的DR病变分割算法,用于同时分割四种不同的DR眼底病变。为了充分利用编码器阶段的多尺度特征信息,在不显著增加网络参数的前提下,提出了非对称式的跳跃连接结构。为了进一步细化特征,并减少特征冗余,在上述结构中添加了注意力模块。在DDR数据集上的实验结果表明,相较于其他DR病变分割方法,本文的算法在保持最小参数量和最快速度的同时,实现了具有高度竞争力的分割性能。

Retinopathy (DR) in diabetes is the main cause of blindness.Fundus lesions are a clinical manifestation of DR,therefore,accurate identification is crucial for early screening,grading,and monitoring of the disease.In this article,a lightweight DR lesion segmentation algorithm is proposed for simultaneously segmenting four different DR fundus lesions.In order to fully utilize the multi-scale feature information of the encoder stage,an asymmetric skip connection structure is proposed without significantly increasing network parameters.In order to further refine the features and reduce feature redundancy,attention modules have been added to the above structure.The experimental results on the DDR dataset show that compared to other DR lesion segmentation methods,our algorithm achieves highly competitive segmentation performance while maintaining the minimum number of parameters and the fastest speed.