基于机器学习的全角膜OCT图像于圆锥角膜早期预警性诊断和预后判断模型研究

Keratoconus(KC) is a binocular corneal ectasia disease, which usually happen at teenager stage without typical signs and symptoms. KC is also one of the several corneal diseases leading to blind. It is important to diagnose KC at early stage, which is helpful for treatment and outcomes. The most widely used corneal topography can only capture the corneal anterior surface changes, which is not sensitive enough for the diagnosing of early KC stage and limited to the experience and subjective judgement from the doctors. The aim of this project is to capture the interfaces and microstructure changes of corneal sub-layers which appear earlier than the corneal anterior surface changes. The entire corneal OCT system can capture the entire corneal sub-layer changes in an non-invasive way. Based on the machine learning method, combined with characteristic indices of the KC OCT images, the early diagnosing and prognosis estimating model will be built. According to the artificial intelligence and computer image processing techniques, the utmost goal of this project is to improve the early detection and prognosis estimating of KC to assist the clinical works.

圆锥角膜是双眼角膜扩张性疾病，多于青壮年发病，早期起病隐匿，无典型临床表现，是致盲性角膜病之一。临床前期圆锥角膜的准确诊断和预后判断，对指导矫正、治疗方式的选择至关重要。目前临床工作常用的角膜地形图，仅可获取临床期圆锥角膜前表面异常，且由于医师诊断的经验及稳定性等瓶颈存在，易造成漏诊。为获取早于角膜地形图出现异常的角膜亚层和表面结构分布改变，本研究拟采用全角膜OCT系统通过非接触、高速获得全角膜范围亚层和表面结构信息。并采用机器学习的方法，将OCT图像特征和圆锥角膜特征性指标有效融合，建立圆锥角膜临床前期预警诊断和临床期预后判断自反馈模型，有效提高圆锥角膜的早期诊断效率和诊断率。本项目旨在基于人工智能和OCT图像识别等计算机技术，为临圆锥角膜床前期预警性诊断和临床期预后判断，提供更为客观、精确的量化参考指标和标准，具有一定的理论意义和学术价值。

圆锥角膜是双眼角膜扩张性疾病，多于青壮年发病，早期起病隐匿，无典型临床表现，是致盲性角膜病之一。临床前期圆锥角膜的准确诊断和预后判断，对指导矫正、治疗方式的选择至关重要。为获取早于角膜地形图出现异常的角膜亚层和表面结构分布改变，本研究采用全角膜地形图获得全角膜范围图像信息。研究共纳入1108例患者1108只眼，其中正常对照组430只眼，亚临床（AKC）患者231只眼，圆锥角膜(KC)患者447只眼。所有患者被分为训练集（664眼）、测试集（222眼）和验证集（222眼）。基于客观指标的全角膜原始数据(KerNet)建立人工智能模型，并对人工智能模型的判别性能进行了评价。人工智能模型的识别能力由准确率和ROC曲线下面积(AUC)评估。KerNet模型在测试集(准确率=94.67%，AUC=0.985)和验证集(准确率=94.12%，AUC=0.990)中具有较强的整体判别能力。在测试集中，KerNet模型对AKC组表现出良好的诊断能力(准确率=95.24%，AUC=0.984)。 验证集也证明KerNet模型对AKC组的诊断是有用的(准确率=94.12%，AUC=0.983)。 基于全角膜地形图原始数据的KerNet人工智能模型，可有效将亚临床圆锥角膜与正常眼区分开来，实现圆锥角膜早期诊断。