Ye Xin, Qiu Wangli, Tu Linzhen, Shen Yingjiao, Chen Qian, Lin Xiangrui, Wang Yaqi, Xie Wenbin, Shen Li-Jun
Zhejiang Provincial People's Hospital, Hangzhou, Zhejiang, China.
Zhejiang Provincial People' s Hospital Bijie Hospital, Bijie, Guizhou, China.
BMJ Open Ophthalmol. 2025 Apr 15;10(1):e002037. doi: 10.1136/bmjophth-2024-002037.
To develop an artificial intelligence (AI) system for detecting pathological patterns of diabetic macular oedema (DME) with fine-grained image categorisation using optical coherence tomography (OCT) images.
The development of the AI system consists of two parts, a pretraining process on a public dataset (Asia Pacific Tele-Ophthalmology Society (APTOS)), and the training process on the local dataset. The local dataset was partitioned into the training set, validation set and test set in the ratio of 6:2:2. The Split Subspace Attention Network (SSA-Net) architecture was adopted to train independent models to detect the seven pathological patterns of DME: intraretinal fluid (IRF), subretinal fluid (SRF), pigment epithelial detachment (PED), hyper-reflective retinal foci (HRF), Müller cell cone disruption (MCCD), subretinal hyper-reflective material (SHRM) and intra-cystic hyper-reflective material (ICHRM). The confusion matrix, sensitivity, specificity and receiver operating characteristic (ROC) were used to evaluate the performance of the models.
The APTOS public dataset consists of 33 853 OCT images and the local dataset consists of 1346 OCT images with DME. In the pretraining process on the public dataset, the accuracy was 0.652 for IRF, 0.928 for SRF, 0.936 for PED and 0.975 for HRF. After the training process on the local dataset, the SSA-Net architecture showed better performance in fine-grained image categorisation on the test set. The area under the ROC curve was 0.980 for IRF, 0.995 for SRF, 0.999 for PED, 0.908 for MCCD, 0.887 for HRF, 0.990 for SHRM and 0.972 for ICHRM. The AI system outputs a heatmap for each result, which can give a visual explanation for the automated identification process. The heatmaps revealed that the regions of interest of the AI system were consistent with the retinal experts.
The proposed SSA-Net architecture for detecting the pathological patterns of DME achieved satisfactory accuracy. However, this study was conducted in one medical centre, and multicentre trials will be needed in the future.
开发一种人工智能（AI）系统，用于通过光学相干断层扫描（OCT）图像的细粒度图像分类来检测糖尿病性黄斑水肿（DME）的病理模式。
AI系统的开发包括两个部分，在公共数据集（亚太远程眼科学会（APTOS））上的预训练过程，以及在本地数据集上的训练过程。本地数据集按6:2:2的比例划分为训练集、验证集和测试集。采用分裂子空间注意力网络（SSA-Net）架构训练独立模型，以检测DME的七种病理模式：视网膜内液（IRF）、视网膜下液（SRF）、色素上皮脱离（PED）、高反射性视网膜病灶（HRF）、米勒细胞锥体破坏（MCCD）、视网膜下高反射物质（SHRM）和囊内高反射物质（ICHRM）。使用混淆矩阵、敏感性、特异性和受试者工作特征（ROC）来评估模型的性能。
APTOS公共数据集由33853张OCT图像组成，本地数据集由1346张患有DME的OCT图像组成。在公共数据集上的预训练过程中，IRF的准确率为0.652，SRF为0.928，PED为0.936，HRF为0.975。在本地数据集上进行训练后，SSA-Net架构在测试集的细粒度图像分类中表现出更好的性能。IRF的ROC曲线下面积为0.980，SRF为0.995，PED为0.999，MCCD为0.908，HRF为0.887，SHRM为0.990，ICHRM为0.972。AI系统为每个结果输出一个热图，可为自动识别过程提供直观解释。热图显示AI系统的感兴趣区域与视网膜专家一致。
所提出的用于检测DME病理模式的SSA-Net架构取得了令人满意的准确率。然而，本研究是在一个医疗中心进行的，未来需要进行多中心试验。
