Ogawa Ryo, Nishikawa Jun, Hideura Eizaburo, Goto Atsushi, Koto Yurika, Ito Shunsuke, Unno Madoka, Yamaoka Yuko, Kawasato Ryo, Hashimoto Shinichi, Okamoto Takeshi, Ogihara Hiroyuki, Hamamoto Yoshihiko, Sakaida Isao
Department of Gastroenterology and Hepatology, Yamaguchi University Graduate School of Medicine, 1-1-1 Minamikogushi, Ube, Japan.
Department of Laboratory Science, Yamaguchi University Graduate School of Medicine, 1-1-1 Minamikogushi, Ube, Japan.
J Gastrointest Cancer. 2019 Sep;50(3):386-391. doi: 10.1007/s12029-018-0083-6.
The utility of chromoendoscopy for early gastric cancer (GC) was determined by machine learning using data of color differences.
Eighteen histopathologically confirmed early GC lesions were examined. We prepared images from white light endoscopy (WL), indigo carmine (Indigo), and acetic acid-indigo carmine chromoendoscopy (AIM). A border between cancerous and non-cancerous areas on endoscopic images was established from post-treatment pathological findings, and 2000 pixels with equivalent luminance values were randomly extracted from each image of cancerous and non-cancerous areas. Each pixel was represented as a three-dimensional vector with RGB values and defined as a sample. We evaluated the Mahalanobis distance using RGB values, indicative of color differences between cancerous and non-cancerous areas. We then conducted diagnosis test using a support vector machine (SVM) for each image. SVM was trained using the 100 training samples per class and determined which area each of 1900 test samples per class came from.
The means of the Mahalanobis distances for WL, Indigo, and AIM were 1.52, 1.32, and 2.53, respectively and there were no significant differences in the three modalities. Diagnosability per endoscopy technique was assessed using the F1 measure. The means of F1 measures for WL, Indigo, and AIM were 0.636, 0.618, and 0.687, respectively. AIM images were better than WL and Indigo images for the diagnosis of GC.
Objective assessment by SVM found AIM to be suitable for diagnosis of early GC based on color differences.
通过使用色差数据的机器学习来确定色素内镜检查对早期胃癌（GC）的效用。
对18个经组织病理学确诊的早期GC病变进行检查。我们准备了白光内镜检查（WL）、靛胭脂（Indigo）和醋酸-靛胭脂色素内镜检查（AIM）的图像。根据治疗后的病理结果确定内镜图像上癌区和非癌区的边界，并从癌区和非癌区的每张图像中随机提取2000个具有等效亮度值的像素。每个像素用RGB值表示为三维向量并定义为一个样本。我们使用RGB值评估马氏距离，其表示癌区和非癌区之间的色差。然后我们对每张图像使用支持向量机（SVM）进行诊断测试。SVM使用每类100个训练样本进行训练，并确定每类1900个测试样本分别来自哪个区域。
WL、Indigo和AIM的马氏距离均值分别为1.52、1.32和2.53，这三种模式之间无显著差异。使用F1测量评估每种内镜检查技术的可诊断性。WL、Indigo和AIM的F1测量均值分别为0.636、0.618和0.687。AIM图像在GC诊断方面优于WL和Indigo图像。
通过SVM进行的客观评估发现AIM适用于基于色差诊断早期GC。
