【摘要】 青光眼已成为全球致盲的主要原因之一。通常,眼科医生利用彩色眼底图像对患者的视神经头(ONH)区域进行评估以诊断青光眼。然而,作为ONH评估重要指标之一的杯盘比(CDR)大都由医生进行人工测量和计算,耗时、费力且带有一定的主观性。为此,提出一种基于改进U-Net的青光眼视杯盘分割算法,在U-Net的编码部分采用ResNet50的映射叠加方式,有效提取图像深层信息。结果表明,所设计模型在公开的DRIONSDB、RIM-ONE和DRISHTI-GS数据集上分别获得AUC值为0.982、0.962和0.989;针对视盘区域分割,IOU分别为0.93、0.94和0.93,Dice系数分别为0.96、0.97和0.97;在RIM-ONE和DRISHTI-GS数据集上,针对视杯区域分割,IOU分别为0.845与0.93,Dice系数分别为0.923与0.967。与眼科医生分割结果相比,其标准误差小于0.16,验证了该算法的优越性能。更多还原

【Abstract】 Glaucoma is one of the leading causes of blindness worldwide. In general,ophthalmologists use color fundus images to evaluate the patient’s optic nerve hypoplasia(ONH)to diagnose glaucoma. Nevertheless,Cup-to-Disc ratio(CDR)as an important indicator of ONH,is mostly measured and calculated by ophthalmologists,which is time-consuming,laborious and subjective. To this end,an end-to-end deep learning approach for optic disc(OD)and optic cup(OC)segmentation in detecting glaucoma based on improved U-Net architecture was presented,which utilized ResNet50 to effectively extract deep feature information on fundus images and boosted its generalization. Further,the ellipse fitting strategy was adopted to improve the edge information of the OD and OC region so that the sawtooth phenomenon could be eliminated in the segmentation prediction. The experimental results showed that on the publicly available DRIONS-DB,RIM-ONE and DRISHTI-GS datasets,our model achieved the AUCs of 0.982,0.962 and 0.989,respectively. The IOU scores of our model for OD segmentation on these datasets were 0.93,0.94 and 0.93,respectively,while the dice coefficients were 0.96,0.97 and 0.97,respectively. In the case of OC segmentation on RIM-ONE and DRISHTI-GS datasets,our model generated the IOUs of 0.845 and 0.93,respectively,whereas obtained 0.923 and 0.967 dice coefficient,respectively. In addition,The model manifested competitive performance(error ranging within 0.16)by contrast to ophthalmologists.更多还原