基于图像与自然语言识别的多模态角膜病人工智能诊断系统的构建与验证

适合医疗场景应用的人工智能（AI）面临的一大实际问题是数据类型繁多、采集难、打标难。目前临床上已经有应用于白内障、视网膜疾病及青光眼诊断的AI系统，遗憾的是AI在角膜病诊断方面的应用研究较少。我们在前期研究工作中基于图像识别构建的眼科AI获得2018年中国国际大数据产业博览会最高奖。目前，该技术可用于常见致盲性眼病的初步分类，并实现感染性与非感染性角膜病的智能诊断，但其尚不能对次级角膜病，如病毒性、真菌性角膜炎、干眼等进行区分。因此，本课题组拟利用图像识别与自然语言识别技术并使用多模型的知识迁移算法，构建基于病史、医学影像及基因测序大数据的多模态角膜病诊断系统，实现人工智能诊断常见角膜病准确率高于90%，假阴性率低于5%，并将其在多家医疗单位中加以应用及验证。本课题组提出的构建多模态AI理论为发展现有眼科AI技术提供了新的解决思路，有利于提升眼科医师的工作效率，并节省时间及人力成本。

The major concerns of artificial intelligence (AI) for medical diagnosis are the variety of data types, the difficulty to collect the data, and the difficulty to mark the images. Currently, there are AI systems used in the diagnosis of cataract, retinal diseases and glaucoma in clinic, unfortunately, studies on the diagnosis of corneal diseases are limited. In our previous project, Deepeye, a disease prediction method based on biological big Data, was developed mainly based on clinical images and won the highest award of the 2018 China International Big Data Industry Expo. So far, this technique can be used for the preliminary classification of common blinding eye diseases, and diagnozing four common corneal diseases (infectious corneal disease, immune corneal disease, corneal degeneration and dystrophy, ocular surface neoplasm), however, it can not be used for distinguishing secondary corneal diseases, such as viral, fungal keratitis, dry eye and so on. Therefore, our group intends to use computer vision and natural language procressing technology and use multi-model knowledge transfer learning algorithm to develop a corneal disease screening and diagnosis system based on medical history, medical imaging and gene sequencing big data. Our goad is to achieve the diagnosis accuracy more than 90%, and the false negative rate not higher than 5%. We aim to apply it to clinical practice in more than 10 hospitals with more than 10000 patients. This artificial intelligence diagnostic systemen would improve the work efficiency and save time and labor costs for ophthalmologists.