目的构建基于深度学习的角膜活体共聚焦显微镜(IVCM)图像人工智能(AI)辅助识别系统，并评估其在临床上的应用价值。
方法采用诊断试验研究方法，收集2021年5月至2022年9月于武汉大学人民医院及武汉大学中南医院行IVCM检查的331例受试者18 860张角膜IVCM图像，经资深角膜专家对图像进行筛选分类后用于模型的训练及测试。模型包括低质量图像过滤模型、角膜图像诊断模型及角膜上皮层、前弹力层、基质层和内皮层4个层次识别模型，判断正常及异常角膜图像与对应角膜层次。选取360张独立于数据库的IVCM图像进行人机大赛，比较3位角膜专业高年资医师与人工智能对图片识别的准确度及用时。另选取8位未经过IVCM培训且临床经验少于3年的低年资医师对360张图像进行无模型辅助及模型辅助下识图，对2次评估结果进行差异分析以评估模型辅助效果。
结果本诊断模型筛选高质量图像的准确度为0.954，在内部及外部测试集中，识别角膜正常/异常图像的总体准确度分别为0.916和0.896；在识别正常及异常图像的角膜层次中，内部测试集的准确度分别达到了0.983及0.925，外部测试集分别达到了0.988及0.929。人机大赛中，模型总体的识别准确度为0.878，与3名高年资医师的平均准确度相近，且评估速度远高于高年资医师，约为其300倍。低年资医师经机器辅助后对图像正/异常及层次诊断的总体平均准确度为0.816±0.043，明显高于模型辅助前的0.669±0.061，差异有统计学意义( t＝6.304， P<0.001)。
结论成功构建了一种基于深度学习的IVCM角膜图像辅助识别系统，可以较准确地分辨正常/异常图像并诊断图像对应角膜层次，能够提高临床诊断效率并辅助医师训练学习。
ObjectiveTo construct an artificial intelligence (AI)-assisted system based on deep learning for corneal in vivo confocal microscopy (IVCM) image recognition and to evaluate its value in clinical applications.
MethodsA diagnostic study was conducted.A total of 18 860 corneal images were collected from 331 subjects who underwent IVCM examination at Renmin Hospital of Wuhan University and Zhongnan Hospital of Wuhan University from May 2021 to September 2022.The collected images were used for model training and testing after being reviewed and classified by corneal experts.The model design included a low-quality image filtering model, a corneal image diagnosis model, and a 4-layer identification model for corneal epithelium, Bowman membrane, stroma, and endothelium, to initially determine normal and abnormal corneal images and corresponding corneal layers.A human-machine competition was conducted with another 360 database-independent IVCM images to compare the accuracy and time spent on image recognition by three senior ophthalmologists and the AI system.In addition, 8 trainees without IVCM training and with less than three years of clinical experience were selected to recognize the same 360 images without and with model assistance to analyze the effectiveness of model assistance.This study adhered to the Declaration of Helsinki.The study protocol was approved by the Ethics Committee of Renmin Hospital of Wuhan University (No.WDRY2021-K148).
ResultsThe accuracy of this diagnostic model in screening high-quality images was 0.954.Its overall accuracy in identifying normal/abnormal corneal images was 0.916 and 0.896 in the internal and external test sets, respectively.Its accuracy reached 0.983, 0.925 in the internal test sets and 0.988, 0.929 in the external test sets in identifying corneal layers of normal and abnormal images, respectively.In the human-machine competition, the overall recognition accuracy of the model was 0.878, which was similar to the average accuracy of the three senior physicians and was approximately 300 times faster than the experts in recognition speed.Trainees assisted by the system achieved an accuracy of 0.816±0.043 in identifying corneal layers of normal and abnormal images, which was significantly higher than 0.669±0.061 without model assistance ( t=6.304, P<0.001).
ConclusionsA deep learning-based assistant system for corneal IVCM image recognition is successfully constructed.This system can discriminate normal/abnormal corneal images and diagnose the corresponding corneal layer of the images, which can improve the efficiency of clinical diagnosis and assist doctors in training and learning.
颜瑜琳,蒋维艳,程思敏,等. 基于深度学习的角膜活体共聚焦显微镜图像辅助识别系统的构建及应用[J]. 中华实验眼科杂志,2024,42(02)：129-135.
DOI:10.3760/cma.j.cn115989-20230101-00001版权归中华医学会所有。
未经授权，不得转载、摘编本刊文章，不得使用本刊的版式设计。
除非特别声明，本刊刊出的所有文章不代表中华医学会和本刊编委会的观点。
颜瑜琳：设计试验、实施研究、采集数据、分析/解释数据、文章撰写；蒋维艳：实施研究、采集数据、分析数据；程思敏：实施研究、采集数据；周奕文、于薏：设计试验、实施研究、采集数据；郑碧清：参与人工智能模型搭建及测试；杨燕宁：参与试验设计、文章审阅及定稿
文档即答
智问智答
机器翻译
你好，我可以帮助您更好的了解本文，请向我提问您关注的问题。
