融合注意力机制和空洞卷积的甲状腺眼病区域分割方法  

A method for segmenting thyroid orbitopathy regions via integration of attention mechanism and dilated convolution

作　　者：李煜豫 何宏 陈家毓 周慧芳[2] LI Yuyu;HE Hong;CHEN Jiayu;ZHOU Huifang(School of Health Science and Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;Department of Ophthalmology,Shanghai Ninth People′s Hospital,Shanghai Jiao Tong University School of Medicine,Shanghai 200011,China)

机构地区：[1]上海理工大学健康科学与工程学院,上海200093 [2]上海交通大学医学院附属第九人民医院眼科,上海200011

出　　处：《智能计算机与应用》2024年第12期51-59,共9页Intelligent Computer and Applications

基　　金：国家科技部项目(G2021013008);上海市科学技术委员会项目(18070503000);上海理工大学医工交叉重点项目(1020308405,1022308502)。

摘　　要：针对眼科门诊对甲状腺眼病分级诊断时,存在诊断主观性较强、工作量大、判断不准确且医护资源消耗较大的问题,为此提出一种融合注意力机制和空洞卷积的甲状腺眼病(Thyroid Associated Ophthalmopathy,TAO)患病区域多类别分割方法。该方法以U型网络(U-Net)作为基础,提出了并行多尺度空洞卷积模块并将其引入下采样底部,通过不同尺度感受野提取因下采样而丢失的有效信息,以提升模型的特征提取能力;此外,针对眼睛区域较小且TAO患者眼睑与巩膜病变后距离发生改变的问题,提出了融合注意力机制模块,提高模型在通道和空间上特征提取的有效性,有效捕获眼部图像的高级与低级语义特征,改善眼睑与巩膜边缘处分割效果。为验证本文方法的有效性,将在数据增强后的2369张巩膜和眼睑数据集上进行训练和预测,实验结果表明所提方法的Dice相似系数、平均交并比(mIoU)分别达到了93.33%、87.61%,实现了对甲状腺眼病患病区域的准确分割,可以有效地辅助甲状腺眼病的严重性分级诊断。In order to solve the problems of strong subjectivity,heavy workload,inaccurate judgment and large consumption of medical resources in the graded diagnosis of thyroid eye disease in ophthalmology clinics,a multi-category segmentation method of Thyroid Associated Ophthalmopathy(TAO)disease area integrating attention mechanism and cavitary convolution is proposed.Based on U-Net,a parallel multi-scale dilated convolution module is proposed and introduced to the bottom of downsampling,and the effective information lost due to downsampling is extracted through different scale receptive fields,so as to improve the feature extraction ability of the model.In addition,in order to solve the problem that the eye area is small and the distance between the eyelid and the scleral lesion changes in TAO patients,a fusion attention mechanism module is proposed in the paper to improve the effectiveness of the model in channel and spatial feature extraction,effectively capture the high-level and low-level semantic features of the eye image,and improve the segmentation effect between the eyelid and scleral edge.In order to verify the effectiveness of the proposed method,the similarity coefficient Dice and mean Intersection Union ratio(mIoU)of the proposed method reach 93.33%and 87.61%,respectively,which realizes the accurate segmentation of the affected area of thyroid eye disease,and could effectively assist in the graded diagnosis of the severity of thyroid eye disease.

关 键 词：甲状腺眼病 多类别分割 空洞卷积 注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]