College of Software, Xinjiang University, Urumqi 830046, China.
State Key Laboratory of Ophthalmology, Zhongshan Ophthalmic Center, Sun Yat-Sen University, Guangzhou, 510060 China.
Photodiagnosis Photodyn Ther. 2022 Mar;37:102689. doi: 10.1016/j.pdpdt.2021.102689. Epub 2021 Dec 18.
Keratitis and conjunctivitis are the most common ocular diseases, their symptoms are similar and easy to confuse, however infectious conjunctivitis is highly contagious. If misdiagnosed, it may worsen the disease and pose a threat to public health.This is a preclinical study to propose a method for rapid and accurate screening of keratitis and conjunctivitis by combining tear Raman spectroscopy with deep learning models that may be applied to clinical applications in the future.The tears of 16 cases of keratitis patients, 13 cases of conjunctivitis patients and 46 cases of healthy subjects were collected, and their Raman spectra were compared and analyzed. By adding different decibels of Gaussian white noise to expand the data, the performance of the tear Raman spectra with a large sample size in the deep learning model was discussed. Principal component analysis (PCA), partial least squares (PLS) and maximum correlation minimum redundancy (mRMR) were used for feature extraction. The processed data were imported into convolutional neural network (CNN) and recurrent neural network (RNN) depth models for classification. After the data were enhanced and processed by PLS, the highest classification accuracy of healthy subjects and keratitis patients, healthy subjects and conjunctivitis patients, and keratitis and conjunctivitis patients reached 94.8%, 95.4%, and 92.7%, respectively. The results of this study show that the use of large sample tear Raman spectra data combined with PLS feature extraction and depth learning algorithms may have great potential in clinical screening of keratitis and conjunctivitis.
角膜炎和结膜炎是最常见的眼部疾病，其症状相似且易于混淆，但传染性结膜炎具有高度传染性。如果误诊，可能会使病情恶化，对公众健康构成威胁。这是一项临床前研究，旨在通过结合泪液拉曼光谱和深度学习模型，提出一种快速准确筛选角膜炎和结膜炎的方法，该方法未来可能应用于临床。收集了 16 例角膜炎患者、13 例结膜炎患者和 46 例健康受试者的泪液，比较和分析了它们的拉曼光谱。通过添加不同分贝的高斯白噪声来扩展数据，讨论了在深度学习模型中具有大样本量的泪液拉曼光谱的性能。采用主成分分析（PCA）、偏最小二乘法（PLS）和最大相关最小冗余度（mRMR）进行特征提取。处理后的数据被导入卷积神经网络（CNN）和循环神经网络（RNN）深度模型进行分类。对 PLS 增强和处理后的数据进行分析，健康受试者和角膜炎患者、健康受试者和结膜炎患者、角膜炎和结膜炎患者的最高分类准确率分别达到 94.8%、95.4%和 92.7%。本研究结果表明，使用大样本量的泪液拉曼光谱数据结合 PLS 特征提取和深度学习算法，在角膜炎和结膜炎的临床筛查中可能具有很大的潜力。
