角膜节段断层扫描生物标志物的通用架构用于人工智能驱动的早期圆锥角膜诊断
本研究利用人工智能对屈光手术患者的节段断层扫描(Placido和光学相干断层扫描)数据进行三维分析,采用Zernike多项式和随机森林模型。在健康眼、极高不对称扩张眼和圆锥角膜眼中,AI分类曲线下面积分别为0.994、0.976和0.997,准确率分别为95.6%、95.5%和99.1%。AI将36只极高不对称扩张眼重新分类为健康(亚临床),104只保留为顿挫型。
要点
- AI模型对健康眼和圆锥角膜眼的分类曲线下面积分别为0.994和0.997
- 极高不对称扩张眼的曲线下面积为0.976,准确率95.5%,召回率71.5%
- AI将36只极高不对称扩张眼重分类为健康(亚临床),104只为顿挫型
核心结论
- 开发了结合节段断层扫描、Zernike多项式和人工智能的通用架构,对健康眼和圆锥角膜眼实现了极好的分类。
- AI有效将极高不对称扩张眼分类为「亚临床」和「顿挫型」。
研究设计与样本
- 使用屈光手术患者的术前影像数据(MS-39,CSO,意大利),患者具有稳定结局并诊断为不对称或双侧圆锥角膜。
- 训练和交叉验证组包括健康眼(n=527)、极高不对称扩张眼(VAE;n=144)和圆锥角膜眼(n=454)。
- 极高不对称扩张眼为圆锥角膜患者的对侧眼,但未进一步手动区分为亚临床或顿挫型。
主要发现
- 健康眼:曲线下面积0.994,准确率95.6%,召回率98.5%,精确率92.7%。
- 圆锥角膜眼:曲线下面积0.997,准确率99.1%,召回率98.7%,精确率99.1%。
- 极高不对称扩张眼:曲线下面积0.976,准确率95.5%,召回率71.5%,精确率91.2%。
- AI将36只(亚临床)极高不对称扩张眼重新分类为健康,尽管这些眼与健康眼不同。
- 大多数剩余的极高不对称扩张眼(n=104;顿挫型)保留其分类,且与圆锥角膜眼和健康眼均不同。
关键数值与推荐
- 健康眼分类:曲线下面积0.994,准确率95.6%,召回率98.5%,精确率92.7%。
- 圆锥角膜眼分类:曲线下面积0.997,准确率99.1%,召回率98.7%,精确率99.1%。
- 极高不对称扩张眼分类:曲线下面积0.976,准确率95.5%,召回率71.5%,精确率91.2%。
临床怎么用
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
AIMS: To develop a comprehensive three-dimensional analyses of segmental tomography (placido and optical coherence tomography) using artificial intelligence (AI). METHODS: Preoperative imaging data (MS-39, CSO, Italy) of refractive surgery patients with stable outcomes and diagnosed with asymmetric or bilateral keratoconus (KC) were used. The curvature, wavefront aberrations and thickness distributions were analysed with Zernike polynomials (ZP) and a random forest (RF) AI model. For training and cross-validation, there were groups of healthy (n=527), very asymmetric ectasia (VAE; n=144) and KC (n=454). The VAE eyes were the fellow eyes of KC patients but no further manual segregation of these eyes into subclinical or forme-fruste was performed. RESULTS: The AI achieved an excellent area under the curve (0.994), accuracy (95.6%), recall (98.5%) and precision (92.7%) for the healthy eyes. For the KC eyes, the same were 0.997, 99.1%, 98.7% and 99.1%, respectively. For the VAE eyes, the same were 0.976, 95.5%, 71.5% and 91.2%, respectively. Interestingly, the AI reclassified 36 (subclinical) of the VAE eyes as healthy though these eyes were distinct from healthy eyes. Most of the remaining VAE (n=104; forme fruste) eyes retained their classification, and were distinct from both KC and healthy eyes. Further, the posterior surface features were not among the highest ranked variables by the AI model. CONCLUSIONS: A universal architecture of combining segmental tomography with ZP and AI was developed. It achieved an excellent classification of healthy and KC eyes. The AI efficiently classified the VAE eyes as 'subclinical' and 'forme-fruste'.
引用
- Kundu G, Shetty R, Khamar P, et al. Universal architecture of corneal segmental tomography biomarkers for artificial intelligence-driven diagnosis of early keratoconus. Br J Ophthalmol. 2021. PMID: 34916211. doi:10.1136/bjophthalmol-2021-319309
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。