Multiclass classification of autofluorescence images of oral cavity lesions based on quantitative analysis
Multiclass classification of autofluorescence images of oral cavity lesions based on quantitative analysis
复制标题
DOI：
10.1371/journal.pone.0228132
复制
发表时间：
2020-02-04
期刊：
影响因子：
3.7
通讯作者：
Chow, Lee
中科院分区：
文献类型：
--
作者：
Jeng, Ming-Jer;Sharma, Mukta;Chow, Lee
背景口腔癌是全球最常见的疾病之一。常规口腔检查和组织病理学检查是临床早期诊断口腔癌的两种主要方法。 VELscope 是一种利用自发荧光的口腔癌筛查设备。当用于区分正常、癌前和恶性病变时，它会产生不一致的结果。我们开发了一种新方法来提高分化的准确性。材料和方法在白光和自发荧光（VELscope，400-460 nm波长）下收集21个正常粘膜以及31个舌和颊粘膜癌前病变和16个恶性病变各5个样本（图像）。使用 iPod（Apple，亚特兰大佐治亚州，美国）开发图像。结果计算归一化强度和强度标准差，以对感兴趣区域 (ROI) 中的图像像素进行分类。使用线性判别分析（LDA）和二次判别分析（QDA）分类器。两个分类器的性能均根据准确度、精确度和召回率进行评估。这些参数用于多类分类。对于舌和颊粘膜，LDA 与非标准化数据的准确率分别提高了 2% 和 14%，QDA 的准确率分别提高了 16% 和 25%。结论 QDA 算法在所有标准评估参数的自发荧光图像分析中均优于 LDA 分类器。
BackgroundOral cancer is one of the most common diseases globally. Conventional oral examination and histopathological examination are the two main clinical methods for diagnosing oral cancer early. VELscope is an oral cancer-screening device that exploited autofluorescence. It yields inconsistent results when used to differentiate between normal, premalignant and malignant lesions. We develop a new method to increase the accuracy of differentiation.Materials and methodsFive samples (images) of each of 21 normal mucosae, as well as 31 premalignant and 16 malignant lesions of the tongue and buccal mucosa were collected under both white light and autofluorescence (VELscope, 400-460 nm wavelength). The images were developed using an iPod (Apple, Atlanta Georgia, USA).ResultsThe normalized intensity and standard deviation of intensity were calculated to classify image pixels from the region of interest (ROI). Linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) classifiers were used. The performance of both of the classifiers was evaluated with respect to accuracy, precision, and recall. These parameters were used for multiclass classification. The accuracy rate of LDA with un-normalized data was increased by 2% and 14% and that of QDA was increased by 16% and 25% for the tongue and buccal mucosa, respectively.ConclusionThe QDA algorithm outperforms the LDA classifier in the analysis of autofluorescence images with respect to all of the standard evaluation parameters.