文献速递神经眼科眼底病与葡萄膜炎

多模态深度学习区分视乳头水肿和非动脉炎性前部缺血性视神经病变与健康眼

Multimodal Deep Learning Differentiates Papilledema and Non-Arteritic Anterior Ischemic Optic Neuropathy From Healthy Eyes.
PubMed 文献 · Invest Ophthalmol Vis Sci · 2026-01-01

本研究开发了基于OCT扫描(3D-ResNet-18)和眼底照片(ResNet-50)的两个深度学习模型,用于分类视乳头水肿(IIH相关)、NAION和健康眼。多模态融合模型(F1WS和XGBoost)的测试准确率分别达到97.5%和98.3%,高于单一OCT模型(93.5%)和眼底照片模型(93.9%)。结论认为结合OCT和眼底照片可提高分类准确性,支持深度学习辅助视乳头水肿的诊断和临床管理。

要点

  • 多模态深度学习融合OCT和眼底照片可高精度区分IIH、NAION和健康眼。
  • 融合模型测试准确率(97.5%和98.3%)优于单一模态模型(93.5%和93.9%)。
  • 该方法有望扩展至黄斑扫描或视野检查等额外数据源。

核心结论

  • 结合OCT和眼底照片的多模态深度学习模型能提高对IIH、NAION和健康眼的分类准确性。

研究设计与样本

  • 开发两个独立模型:OCT扫描(7019张,3D-ResNet-18)和眼底照片(17,657张,ResNet-50),分类视乳头水肿(2315 OCT,6349眼底)、NAION(841 OCT,1814眼底)和健康眼(3863 OCT,9494眼底)。
  • 测试集完全由同日OCT扫描和眼底照片组成,每种模态(OCT和眼底)为每个类别贡献至少15%的数据。
  • 通过两种方法结合两个模型的输出概率:按类F1加权求和(F1WS)和XGBoost模型。

主要发现

  • OCT模型单独测试准确率为93.5%,眼底照片模型为93.9%。
  • 多模态F1WS和XGBoost模型准确率分别为97.5%和98.3%。

临床怎么用

  • 多模态深度学习可辅助视乳头水肿的诊断和临床管理。
  • 可考虑整合OCT和眼底照片两种影像模态以提高诊断准确性。

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

PURPOSE: Optic nerve head (ONH) swelling, a critical feature in idiopathic intracranial hypertension (IIH) and non-arteritic anterior ischemic optic neuropathy (NAION), can present diagnostic challenges. We explored a multimodal deep learning (DL) approach integrating optical coherence tomography (OCT) scans and fundus photographs to enhance diagnostic accuracy for differentiating IIH, NAION, and healthy eyes. METHODS: We developed two separate models using 7019 OCT scans (3D-ResNet-18) and 17,657 fundus photos (ResNet-50) to classify eyes with papilledema (2315 OCT, 6349 fundus), NAION (841 OCT, 1814 fundus), and healthy eyes (3863 OCT, 9494 fundus). We arranged the dataset so that the test set consisted entirely of same-day OCT scans and fundus photos, with each modality (OCT and fundus) contributing at least 15% of the data for each class. We combined output probabilities from both models using two methods: an F1-weighted sum by class (F1WS), as well as an XGBoost model. Performance of each was evaluated with AUC-ROC, accuracy, precision, recall, and F1 scores. RESULTS: The OCT model alone achieved a test accuracy of 93.5%, with the fundus photo model reaching 93.9%. The multimodal F1WS and XGBoost models achieved an accuracy of 97.5% and 98.3%, respectively. CONCLUSIONS: Combining OCT and fundus photographs improves the classification of IIH, NAION, and healthy eyes, showing the value of using complementary imaging modalities. This approach supports the use of DL to aid diagnosis and clinical management of optic nerve head swelling. It may also be extended to leverage DL from additional data sources, such as macular scans or visual field tests.

来源:Invest Ophthalmol Vis Sci,查看原文。版权归原作者所有。

引用

  1. Szanto D, Erekat A, Woods B, et al. Multimodal Deep Learning Differentiates Papilledema and Non-Arteritic Anterior Ischemic Optic Neuropathy From Healthy Eyes. Invest Ophthalmol Vis Sci. 2026. PMID: 41533910. doi:10.1167/iovs.67.1.12

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。