文献速递眼底病与葡萄膜炎

利用卷积神经网络预测糖尿病性黄斑水肿对玻璃体腔抗VEGF治疗的反应:一项初步研究

Use of a Convolutional Neural Network to Predict the Response of Diabetic Macular Edema to Intravitreal Anti-VEGF Treatment: A Pilot Study.
PubMed 文献 · Am J Ophthalmol · 2025-02-20

该单中心回顾性研究纳入53例患者的73只眼,利用基线OCT扫描训练卷积神经网络(CNN)预测初治糖尿病性黄斑水肿(DME)对单次玻璃体腔抗血管内皮生长因子(anti-VEGF)注射的反应。以中心子区域厚度(CST)下降≥10 µm定义治疗反应,57只眼为有反应者、16只眼为无反应者;CNN的曲线下面积(AUC)为0.81,召回率78.9%,特异度68.8%。作者认为结果展示了CNN预测初治DME对单次anti-VEGF治疗反应的潜力。

要点

  • CNN仅用基线OCT扫描预测初治DME对单次anti-VEGF注射的反应,AUC为0.81。
  • 以CST下降≥10 µm定义治疗反应,57只眼有反应、16只眼无反应。
  • 在报告的工作点,CNN正确识别57只有反应眼中的45只(召回率78.9%),正确识别16只无反应眼中的11只(特异度68.8%)。

核心结论

  • 作者认为,本研究结果展示了CNN预测初治DME对单次anti-VEGF治疗反应的潜力。

研究设计与样本

  • 本研究为通过病历回顾进行的回顾性研究,在南卡罗来纳医科大学Storm Eye Institute单中心开展。
  • 纳入新诊断DME并接受玻璃体腔anti-VEGF注射、且有诊断时基线OCT和首次注射后1个月随访OCT的患者,共纳入53例患者的73只眼。
  • 排除标准包括既往接受过anti-VEGF治疗、缺少所需OCT扫描、合并黄斑变性和其他视网膜疾病所致黄斑水肿。
  • 将基线OCT与首次anti-VEGF注射后约1个月的随访OCT比较,确定中心子区域厚度变化(delta CST),并将其作为标签训练CNN仅从基线OCT预测治疗反应。

主要发现

  • 根据2次OCT的delta CST,57只眼为首次anti-VEGF注射的有反应者,16只眼为无反应者。
  • 仅分析每只眼的基线OCT扫描时,训练后的CNN曲线下面积(AUC)为0.81。
  • 在报告的工作点,CNN正确识别57只有反应眼中的45只(召回率78.9%),正确识别16只无反应眼中的11只(特异度68.8%)。

关键数值与推荐

  • 治疗反应定义为CST下降10 µm或以上。
  • CNN预测治疗反应的AUC为0.81,召回率为78.9%,特异度为68.8%。

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

PURPOSE: To utilize a convolutional neural network (CNN) to predict the response of treatment-naïve diabetic macular edema (DME) to a single injection of anti-vascular endothelial growth factor (anti-VEGF) with data from optical coherence tomography (OCT). DESIGN: Retrospective study performed via chart review. METHODS: Setting: This was a single-center study performed at the Storm Eye Institute, Medical University of South Carolina. PATIENT POPULATION: Patients with a new diagnosis of DME who underwent intravitreal (IVT) anti-VEGF injections were eligible for inclusion, provided they had a baseline OCT scan at the time of diagnosis and a 1-month follow-up OCT scan after the first anti-VEGF injection. Exclusion criteria included prior treatment with anti-VEGF, lack of required OCT scans, coexistent macular degeneration, and macular edema due to other retinal diseases. Seventy-three (73) eyes from 53 patients were included. INTERVENTION: The OCT scan from the baseline visit was compared to the follow-up OCT scan approximately 1 month after the first anti-VEGF injection to determine change in central subfield thickness (delta CST). The delta CST was fed into the CNN as a label to train the system to predict treatment response from only the baseline OCT scan. MAIN OUTCOME MEASURE: CNN prediction of treatment response to anti-VEGF. Treatment response was defined as a CST reduction of 10 µm or more. RESULTS: Based on delta CST from 2 OCT scans, 57 eyes were responders and 16 eyes were non-responders to the initial anti-VEGF injection. Analyzing only the baseline OCT scan for each eye, the trained CNN demonstrated an area under the curve (AUC) of 0.81. At the reported operating point, the CNN correctly identified 45 of the 57 responder eyes (i.e., recall of 78.9%) and 11 of the 16 non-responder eyes (i.e., specificity of 68.8%). CONCLUSIONS: The results of this study demonstrate the potential of a CNN to predict the response of treatment-naïve DME to a single injection of anti-VEGF therapy. NOTE: Publication of this article is sponsored by the American Ophthalmological Society.

来源:Am J Ophthalmol,查看原文。版权归原作者所有。

引用

  1. Magrath G, Luvisi J, Russakoff D, et al. Use of a Convolutional Neural Network to Predict the Response of Diabetic Macular Edema to Intravitreal Anti-VEGF Treatment: A Pilot Study. Am J Ophthalmol. 2025. PMID: 39986639. doi:10.1016/j.ajo.2025.02.017

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。