Toward automated severe pharyngitis detection with smartphone camera using deep learning networks

Here we present a deep learning model with smartphone-based throat images facilitating detection of severe pharyngitis in telemedicine settings. We collected throat images from the web-based open social Q&A systems including Naver Korea (https://kin.naver.com), Yahoo Japan (https://chiebukuro.yahoo.co.jp). The additional throat image datasets were extracted using the Google image search engine. The search strategy was based on the key terms “sore throat”, “pharyngitis”, “tonsillitis”, “exudative tonsillitis”, “tonsillopharyngitis”, “throat image”, and “smartphone” in Korean, Japanese, and English. The most updated electronic database search was on June 30, 2020. We manually excluded throat images which were not acquired using smartphone. The images with the characteristics of the pharyngitis were manually classified by two clinicians, and the ambiguous images were isolated to clarify the image domains. Finally, we collected the initial dataset with a total of two classes including 147 throat images with pharyngitis and 215 normal throat images.本研究提出一款基于智能手机采集咽喉图像的深度学习模型，可助力远程医疗场景下的重症咽炎检测。 我们从基于网页的开放社交问答系统中采集咽喉图像，其中包括韩国Naver问答平台（https://kin.naver.com）与日本雅虎知惠袋（https://chiebukuro.yahoo.co.jp）。其余咽喉图像数据集则通过谷歌（Google）图片搜索引擎抓取获取。本次搜索以韩语、日语及英语的如下关键词为依据："sore throat"（咽喉痛）、"pharyngitis"（咽炎）、"tonsillitis"（扁桃体炎）、"exudative tonsillitis"（渗出性扁桃体炎）、"tonsillopharyngitis"（扁桃体咽炎）、"throat image"（咽喉图像）以及"smartphone"（智能手机）。最后一次电子数据库检索的时间为2020年6月30日。我们人工剔除了非智能手机采集的咽喉图像。两名临床医师手动标注具有咽炎特征的图像，同时将边界模糊的图像单独分出以明确图像范畴。最终，我们构建了初始数据集，共包含2个类别：147张咽炎咽喉图像与215张正常咽喉图像。