面向AS-OCT的两阶段巩膜突自动定位算法

A two-stage scleral spur automatic localizationalgorithm facing AS-OCT

随着青光眼等前节疾病发病率的增加,眼前节图像的分析受到越来越多的关注。其中,巩膜突(scleral spur,SS)的精确定位对眼前节结构参数的测量和临床诊断具有重要意义。面向眼前节光学相干断层扫描图像提出了一种由“粗”到“细”的两阶段巩膜突自动定位网络。首先,通过一个ROI区域检测网络来识别以巩膜突为中心的ROI区域,在骨干网络中加入ACmix模块,提升了网络的特征提取能力。然后使用轻量级网络Lite-HRNet作为微调修正网络进一步提取ROI图像细节特征,基于热图回归得到更加精确的巩膜突位置。最后通过在两个数据集进行验证,分别将巩膜突定位平均误差降至11.74和9.82。实验证明该方法优于现有其他基于深度学习的方法,具有良好的定位效果。

With the increasing incidence of anterior segment diseases such as glaucoma,the analysis of anterior segment images has received more and more attention.Moreover,the precise localization of the scleral spur(SS)is of great significance for the measurement of anterior segment structural parameters and clinical diagnosis.It presents a two-stage localization network from"coarse"to"fine"for optical coherence tomography images of the anterior segment.First,a ROI region detection network was used to identify ROI regions centered on the scleral spur,and the feature extraction ability of the network was improved by adding the ACmix module to the backbone network.Afterwards,a lightweight network Lite-HRNet was used as a fine-tuning correction network to further extract image detail features,based on heatmap regression to obtain a more accurate scleral process location.Finally,we tested our method on two datasets,reaching an average error of 11.74 and 9.82 respectively.Many experiments show that our method is better than other existing deep learning-based methods and has good detection effect.