Shiihara Hideki, Sonoda Shozo, Terasaki Hiroto, Kakiuchi Naoko, Shinohara Yuki, Tomita Masatoshi, Sakamoto Taiji
Department of Ophthalmology, Kagoshima University Graduate School of Medical and Dental Sciences, Kagoshima, Japan.
Jpn J Ophthalmol. 2018 Nov;62(6):643-651. doi: 10.1007/s10384-018-0625-2. Epub 2018 Oct 6.
To develop an automated method to segment the choroidal layers of en face optical coherent tomography (OCT) images by machine learning.
A cross-sectional, prospective study of 276 eyes of 181 healthy subjects.
OCT en face images of the choroid were obtained every 2.6 μm from the retinal pigment epithelium (RPE) to the chorioscleral border. The images at the start of the choriocapillaris, start of Sattler's layer, and start of Haller's layer were identified, and the image numbers from the RPE line were taken as the teacher data. Forty-one feature quantities of each image were extracted. A support vector machine (SVM) model was created from each feature value of the training data, and a coefficient of determination was calculated for each layer of the choroid by a fivefold cross validation. Next, the same evaluation was performed after creating a SVM model with selected effective feature quantities.
The mean coefficient of determination using all features was 0.9853 ± 0.0012. Nine effective feature quantities (relative choroid thickness, mean/kurtosis/variance of brightness, FFT_ skewness, k0_vessel width, k1/k2/k4_vessel area) were selected, and the mean of the coefficient of determinations with these quantities In this model was 0.9865 ± 0.0001. The number of errors in the image number at the start of each layer was 1.01 ± 0.79 for the choriocapillaris, 1.13 ± 1.12 for Sattler's layer, and 3.77 ± 2.90 for Haller's layer.
Automated stratification of the choroid in en face images can be done with high accuracy through machine learning.
开发一种通过机器学习对眼底光学相干断层扫描（OCT）图像中的脉络膜层进行分割的自动化方法。
对181名健康受试者的276只眼睛进行横断面、前瞻性研究。
从视网膜色素上皮（RPE）到脉络膜巩膜边界，每隔2.6μm获取脉络膜的OCT眼底图像。确定脉络膜毛细血管层起始处、萨特勒层起始处和哈勒层起始处的图像，并将从RPE线开始的图像编号作为教师数据。提取每张图像的41个特征量。根据训练数据的每个特征值创建支持向量机（SVM）模型，并通过五重交叉验证计算脉络膜各层的决定系数。接下来，在使用选定的有效特征量创建SVM模型后进行相同的评估。
使用所有特征的平均决定系数为0.9853±0.0012。选择了9个有效特征量（相对脉络膜厚度、亮度均值/峰度/方差、快速傅里叶变换偏度、k0血管宽度、k1/k2/k4血管面积），该模型中这些特征量的决定系数平均值为0.9865±0.0001。脉络膜毛细血管层起始处图像编号的错误数量为1.01±0.79，萨特勒层为1.13±1.12，哈勒层为3.77±2.90。
通过机器学习可以高精度地对眼底图像中的脉络膜进行自动分层。
