文献速递青光眼眼底病与葡萄膜炎

双输入卷积神经网络用于青光眼诊断:基于SD-OCT的评估

Dual-input convolutional neural network for glaucoma diagnosis using spectral-domain optical coherence tomography.
PubMed 文献 · Br J Ophthalmol · 2020-09-12

该研究评估了一种基于SD-OCT的深度学习分类器(双输入卷积神经网络DICNN)的青光眼诊断能力。研究纳入315名正常者、219例早期原发性开角型青光眼(POAG)和243例中重度POAG的777组Cirrus高清SD-OCT图像(RNFL和GCIPL)。在测试集上,DICNN区分青光眼与正常者的准确率为92.793%,AUC为0.957;区分早期青光眼与正常者的准确率为85.185%,AUC为0.869,且优于单独使用RNFL或GCIPL训练的卷积神经网络算法。

要点

  • DICNN以RNFL和GCIPL厚度图双输入,测试集区分青光眼与正常者准确率92.793%,AUC 0.957。
  • 区分早期POAG与正常者准确率85.185%,AUC 0.869,敏感度0.921,特异度0.756。
  • 双输入DICNN对早期青光眼的诊断能力高于单独RNFL或GCIPL训练的CNN算法。

核心结论

  • 基于SD-OCT的深度学习算法不仅能区分确诊青光眼患者与正常者,还能区分早期青光眼患者与正常者。
  • 同时以RNFL和GCIPL厚度图数据训练的双输入卷积神经网络(DICNN)在区分早期青光眼与正常者方面表现出高诊断能力。

研究设计与样本

  • 研究共纳入777组Cirrus高清SD-OCT图像(视网膜神经纤维层RNFL和神经节细胞-内丛状层GCIPL),来自315名正常者、219例早期原发性开角型青光眼(POAG)患者和243例中重度POAG患者。
  • 图像集分为训练集(252名正常、174例早期POAG、195例中重度POAG)和测试集(63名正常、45例早期POAG、48例中重度POAG)。
  • 采用基于视觉几何组(VGG16)的双输入卷积神经网络(DICNN)进行青光眼诊断,该网络以RNFL和GCIPL两张图像作为输入。
  • 根据准确率和受试者工作特征曲线下面积(AUC)计算青光眼诊断能力。

主要发现

  • 在测试集上,DICNN区分青光眼患者与正常者的准确率为92.793%,AUC为0.957(95% CI 0.943至0.966),敏感度为0.896(95% CI 0.896至0.917),特异度为0.952(95% CI 0.921至0.952)。
  • 区分早期青光眼患者与正常者时,DICNN的准确率为85.185%,AUC为0.869(95% CI 0.825至0.879),敏感度为0.921(95% CI 0.813至0.905),特异度为0.756(95% CI 0.610至0.790)。
  • DICNN对早期青光眼的诊断能力高于单独使用RNFL或GCIPL训练的卷积神经网络算法。

关键数值与推荐

  • 区分青光眼与正常者:准确率92.793%,AUC 0.957(95% CI 0.943至0.966),敏感度0.896(95% CI 0.896至0.917),特异度0.952(95% CI 0.921至0.952)。
  • 区分早期青光眼与正常者:准确率85.185%,AUC 0.869(95% CI 0.825至0.879),敏感度0.921(95% CI 0.813至0.905),特异度0.756(95% CI 0.610至0.790)。

临床怎么用

  • 基于SD-OCT的DICNN模型可作为辅助工具,帮助从正常人群中识别早期青光眼患者,提升早期诊断能力。
  • 临床应用中可优先考虑同时输入RNFL和GCIPL厚度图的深度学习模型,而非仅依赖单一图层数据。

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

BACKGROUND/AIMS: To evaluate, with spectral-domain optical coherence tomography (SD-OCT), the glaucoma-diagnostic ability of a deep-learning classifier. METHODS: A total of 777 Cirrus high-definition SD-OCT image sets of the retinal nerve fibre layer (RNFL) and ganglion cell-inner plexiform layer (GCIPL) of 315 normal subjects, 219 patients with early-stage primary open-angle glaucoma (POAG) and 243 patients with moderate-to-severe-stage POAG were aggregated. The image sets were divided into a training data set (252 normal, 174 early POAG and 195 moderate-to-severe POAG) and a test data set (63 normal, 45 early POAG and 48 moderate-to-severe POAG). The visual geometry group (VGG16)-based dual-input convolutional neural network (DICNN) was adopted for the glaucoma diagnoses. Unlike other networks, the DICNN structure takes two images (both RNFL and GCIPL) as inputs. The glaucoma-diagnostic ability was computed according to both accuracy and area under the receiver operating characteristic curve (AUC). RESULTS: For the test data set, DICNN could distinguish between patients with glaucoma and normal subjects accurately (accuracy=92.793%, AUC=0.957 (95% CI 0.943 to 0.966), sensitivity=0.896 (95% CI 0.896 to 0.917), specificity=0.952 (95% CI 0.921 to 0.952)). For distinguishing between patients with early-stage glaucoma and normal subjects, DICNN's diagnostic ability (accuracy=85.185%, AUC=0.869 (95% CI 0.825 to 0.879), sensitivity=0.921 (95% CI 0.813 to 0.905), specificity=0.756 (95% CI 0.610 to 0.790)]) was higher than convolutional neural network algorithms that trained with RNFL or GCIPL separately. CONCLUSION: The deep-learning algorithm using SD-OCT can distinguish normal subjects not only from established patients with glaucoma but also from patients with early-stage glaucoma. The deep-learning model with DICNN, as trained by both RNFL and GCIPL thickness map data, showed a high diagnostic ability for discriminatingpatients with early-stage glaucoma from normal subjects.

来源:Br J Ophthalmol,查看原文。版权归原作者所有。

引用

  1. Sun S, Ha A, Kim YK, et al. Dual-input convolutional neural network for glaucoma diagnosis using spectral-domain optical coherence tomography. Br J Ophthalmol. 2020. PMID: 32920530. doi:10.1136/bjophthalmol-2020-316274

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。