基于大数据深度学习的甲状腺相关眼病辅助诊断算法的研究

Thyroid associated ophthalmopathy (TAO) is a comparatively uncommon autoimmune disease, but the most common orbitopathy. Therapy effect of severe TAO is poor. A well trained orbitopathy specialist can diagnosis TAO in the early stage and make accurate assessment on its activity level and severity degree through basic inquiry and examination, and further provide reasonable treatment according to the EUGOGO guide and help stopping the develop of TAO. However, there are only few orbitopathy professional doctor. TAO is difficult to diagnose and make correct treatment, especially in the early stage for unprofessional doctors. The assessment of its activity level and severity degree is also an unsettled issue. In this study, traditional machine learning technology and deep learning technology, such as Deep Neural Network and Support Vector Machine, will be used to build algorithms detecting hyperaemia of eyelid and conjunctiva, edema of eyelid, conjunctiva and lacrimal caruncle, as well as corneal ulcer. Algorithms that detect eyelid retraction and eye movement disorder will also be designed. Meanwhile, deep learning/machine learning and extra image processing technology will be used to measure the diameter and area of extra ocular muscles and the degree of exophthalmoses. All these models would be trained by using more than 80000 well annotated eye photos and orbital CT images, collecting from more than 1200 patients with TAO and more than 800 patients without TAO in our department since 2013 and nearly 200 healthy volunteers from our university. These algorithm models will be fused together with simple medical history to form a software which can help medical staff to diagnose TAO in early stage, as well as to assess its activity level and severity degree, and finally to reduce the misdiagnosis, missed diagnosis and therapeutic error effectively.

甲状腺相关眼病（TAO）是最常见的眼眶病，属自身免疫疾病，重度TAO治疗效果差。专业的眼眶病医师通过问诊、视诊和基本检查可早期诊断并准确分期、分度、合理治疗。然此专业医师数量极少，且TAO病情复杂，易被非此专业的医师漏诊、误诊、误治，主要是难以早期诊断和准确分期、分度。本研究拟采用深度卷积神经网络等深度学习技术，基于我科5年来1200余例TAO患者和千余例其他患者或志愿者的8万余张外观照和眼眶CT图像，进行分类、检测和分割标注，获得检测眼睑、结膜、角膜、泪阜等相关体征和眼球运动情况的算法模型以及检测眼外肌直径、面积和眼球突出度的算法模型。融合上述图像检测模型，辅以简单病史修正，获得TAO及其活动性、严重度的人工智能诊断算法模型，验证其准确性和特异性，最终形成应用软件，帮助广大社区医务人员早期诊断TAO并准确分期、分度，进一步依据指南合理治疗并预防TAO进展，有效降低误诊、漏诊率，减少误治。

甲状腺相关眼病（TAO）是最常见的眼眶病，属自身免疫疾病，重度TAO治疗效果差。专业的眼眶病医师通过问诊、视诊和基本检查可早期诊断并准确分期、分度、合理治疗。然此专业医师数量极少，且TAO病情复杂，易被非此专业的医师漏诊、误诊、误治，主要是难以早期诊断和准确分期、分度。.本研究采用深度卷积神经网络等深度学习技术和机器学习技术，基于我科1200余例TAO患者和800余例其他患者或志愿者的8万余张外观照和眼眶CT图像，进行分类、检测和分割标注，获得检测眼睑、结膜、角膜、泪阜等相关体征和眼球运动情况的算法模型以及检测眼外肌增粗和眼球突出度的算法模型。.在眼睛位置检测模型中，在阈值IoU>0.5的情况下，眼睛定位模型的准确度达到0.98，几乎可以高精度地检测出正位及左右侧位三种体态的所有眼睛位置。对于角膜/巩膜的语义分割，U-Net分割模型的准确率分别达到0.93和0.87。.通过多个阶段的深度学习模型预测，能够有效地识别7种TAO体征，其准确率（AUC，95%CI）分别为：眼睑退缩 0.93，眼睑水肿 0.90，眼睑充血 0.94，结膜充血 0.91，结膜水肿 0.60，角膜溃疡 0.79，眼球运动障碍 0.91。总平均AUCs>0.85（95%CI），F1 score>0.80。除对于需要其他辅助手段例如裂隙灯的结膜水肿，及本身较为少见的角膜溃疡外，我们开发的TAO多体征诊断系统可以对大多数常见TAO体征进行较高准确率的检测。.模型在区分眼眶 CT 和非 CT 图像方面准确率（AUC，95%CI）为0.99，确定水平位或冠状位图像有无诊断价值准确率（AUC，95%CI）分别为0.92和0.85。但在冠状CT图像评价眼外肌肿大、水平CT图像评价眼外肌肿大和眼球突出的表现并不乐观，准确率（AUC，95%CI）分别为0.67 、0.53和 0.69。实验结果表明，基于二维CT图像的人工智能诊断TAO诊断系统仍需要更多的数据和训练来提高其诊断性能。.融合上述图像检测模型，辅以简单病史修正，即获得TAO及其活动性、严重度的人工智能诊断算法模型，帮助广大社区医务人员早期诊断TAO并准确分期、分度，进一步依据指南合理治疗并预防TAO进展，有效降低误诊、漏诊率，减少误治。