基于深度学习的多模态眼科图像回归预测

青光眼是世界第二大致盲性眼病，视网膜神经纤维层（RNFL）缺损是诊断青光眼的重要特征。在临床应用中主要采用光学相干断层扫描（OCT）测量RNFL厚度。然而在我国的多数中小型医院和体检中心，只有眼底照相机而不具备OCT设备。因此利用眼底照和OCT的多模态数据，设计了一种基于眼底照来预测RNFL厚度的深度残差回归神经网络。该网络通过眼底照中的局部区域信息预测此区域的RNFL厚度，并对视盘外围一周范围内的RNFL厚度给出全面的刻画。在一个来自北京同仁医院的真实数据集上的实验结果显示，本文算法预测的RNFL厚度值与OCT测量值具有高度的一致性（对于正常眼平均绝对误差E_MA=14.884，Pearson相关系数r=0.885，决定系数R²=0.781；对于青光眼E_MA=15.108，r=0.872，R²=0.754）。评估结果表明所提方法对基于眼底照预测RNFL厚度具有良好的临床实用性。

Regression prediction of multimodal ophthalmology images based on deep learning

Glaucoma is the second leading cause of blindness globally. Retinal nerve fiber layer(RNFL) defects are an important indicator for glaucoma diagnosis. Optical coherence tomography (OCT) is a major method for measuring RNFL thickness in clinical applications. However, most of the small and medium-sized hospitals and physical examination centers in China only have fundus cameras and lack OCT. In this work, we present a deep residual regression network based on multimodal data from fundus photography and OCT reports to predict RNFL thickness, which used local information in fundus images. The proposed method has been verified on a dataset from Beijing Tongren Hospital. The experiments showed that we achieved significant consistency (normals: mean absolute error E_MA=14.884, Pearson correlation coefficient r=0.885, coefficient of determination R²=0.781; glaucomatous: E_MA=15.108, r=0.872, R²=0.754) between values predicted by the model output and the RNFL thickness measured by OCT. The evaluation results thus show the good clinical practicability of the proposed approach to predict RNFL thickness based on fundus photography.