Department of Computer Science, Stanford University, Stanford, California.
Department of Biomedical Informatics, Stanford University, Stanford, California.
Ophthalmology. 2022 Feb;129(2):139-146. doi: 10.1016/j.ophtha.2021.07.033. Epub 2021 Aug 2.
To develop and evaluate an automated, portable algorithm to differentiate active corneal ulcers from healed scars using only external photographs.
A convolutional neural network was trained and tested using photographs of corneal ulcers and scars.
De-identified photographs of corneal ulcers were obtained from the Steroids for Corneal Ulcers Trial (SCUT), Mycotic Ulcer Treatment Trial (MUTT), and Byers Eye Institute at Stanford University.
Photographs of corneal ulcers (n = 1313) and scars (n = 1132) from the SCUT and MUTT were used to train a convolutional neural network (CNN). The CNN was tested on 2 different patient populations from eye clinics in India (n = 200) and the Byers Eye Institute at Stanford University (n = 101). Accuracy was evaluated against gold standard clinical classifications. Feature importances for the trained model were visualized using gradient-weighted class activation mapping.
Accuracy of the CNN was assessed via F score. The area under the receiver operating characteristic (ROC) curve (AUC) was used to measure the precision-recall trade-off.
The CNN correctly classified 115 of 123 active ulcers and 65 of 77 scars in patients with corneal ulcer from India (F score, 92.0% [95% confidence interval (CI), 88.2%-95.8%]; sensitivity, 93.5% [95% CI, 89.1%-97.9%]; specificity, 84.42% [95% CI, 79.42%-89.42%]; ROC: AUC, 0.9731). The CNN correctly classified 43 of 55 active ulcers and 42 of 46 scars in patients with corneal ulcers from Northern California (F score, 84.3% [95% CI, 77.2%-91.4%]; sensitivity, 78.2% [95% CI, 67.3%-89.1%]; specificity, 91.3% [95% CI, 85.8%-96.8%]; ROC: AUC, 0.9474). The CNN visualizations correlated with clinically relevant features such as corneal infiltrate, hypopyon, and conjunctival injection.
The CNN classified corneal ulcers and scars with high accuracy and generalized to patient populations outside of its training data. The CNN focused on clinically relevant features when it made a diagnosis. The CNN demonstrated potential as an inexpensive diagnostic approach that may aid triage in communities with limited access to eye care.
开发并评估一种使用外部照片即可区分活动性角膜溃疡和愈合疤痕的自动化、便携式算法。
使用角膜溃疡和疤痕的照片对卷积神经网络进行训练和测试。
从类固醇性角膜溃疡试验（SCUT）、真菌性溃疡治疗试验（MUTT）和斯坦福大学拜尔斯眼科研究所获得角膜溃疡的去识别照片。
使用来自 SCUT 和 MUTT 的角膜溃疡（n=1313）和疤痕（n=1132）的照片对卷积神经网络（CNN）进行训练。该 CNN 在印度眼科诊所的 2 个不同患者群体（n=200）和斯坦福大学拜尔斯眼科研究所（n=101）进行了测试。准确性是通过金标准临床分类来评估的。使用梯度加权类激活映射可视化训练模型的特征重要性。
通过 F 分数评估 CNN 的准确性。接收器操作特征（ROC）曲线下的面积（AUC）用于衡量精度-召回率的权衡。
CNN 在来自印度的角膜溃疡患者中正确分类了 123 例活动性溃疡中的 115 例和 77 例疤痕中的 65 例（F 分数，92.0%[95%置信区间（CI），88.2%-95.8%]；灵敏度，93.5%[95% CI，89.1%-97.9%]；特异性，84.42%[95% CI，79.42%-89.42%]；ROC：AUC，0.9731）。CNN 在来自北加州的角膜溃疡患者中正确分类了 55 例活动性溃疡中的 43 例和 46 例疤痕中的 42 例（F 分数，84.3%[95% CI，77.2%-91.4%]；灵敏度，78.2%[95% CI，67.3%-89.1%]；特异性，91.3%[95% CI，85.8%-96.8%]；ROC：AUC，0.9474）。CNN 的可视化结果与临床相关特征相关，如角膜浸润、前房积脓和结膜充血。
CNN 以高精度分类了角膜溃疡和疤痕，并且可以推广到其训练数据之外的患者群体。CNN 在做出诊断时关注临床相关特征。CNN 具有作为一种廉价诊断方法的潜力，可能有助于在眼保健资源有限的社区进行分诊。
