RETFound与监督卷积神经网络的比较,用于从 Fundus 照片中检测可推的玻璃眼
Kyle Bolo1, Tran Huy Nguyen2, Sreenidhi Iyengar2
1Roski Eye Institute, Keck School of Medicine, University of Southern California, Los Angeles, California.
Ophthalmology science
|January 19, 2026
概括
像RETFound这样的视力变压器基础模型显示出检测可转移的玻璃眼的前景,特别是在有限的数据下. 剪裁图像可以提高性能,但可能会影响概括性.
科学领域:
- 眼科医生 眼科 眼科
- 人工智能的人工智能
- 医疗成像医学成像
背景情况:
- 从 fundus 照片中检测眼对于预防视力损失至关重要.
- 深度学习模型为自动选提供了潜力.
- 对比不同的AI架构对于临床采用至关重要.
研究的目的:
- 将基于视力变压器的基础模型 (RETFound) 与卷积神经网络 (VGG-19) 的诊断性能进行比较,以检测可转移的玻璃眼.
- 为了评估各种培训集大小和人口统计因素的模型性能.
主要方法:
- 在剪切和未剪切的基底图像上训练了四个深度学习模型 (RETFound和VGG-19).
- 利用了来自洛杉矶县远视膜查计划的大型数据集.
- 在使用AUC-ROC和值特定指标的持有内部和外部测试集上评估模型.
主要成果:
- 与未经裁剪的模型相比,切割的图像模型 (VGG-19和RETFound) 显示出更高的性能.
- 当在较小的数据集上进行训练时,RETFound模型表现出了优势,并在各族群中显示出一致的性能.
- 没有切割的RETFound模型在外部验证数据上表现最好.
结论:
- 无论是RETFound还是VGG-19模型都能有效地检测可转移的玻璃眼.
- 对于有限的培训数据和预期的领域转移,像RETFound这样的基础模型可能更好.
- 剪裁图像可以提高性能,但可能会降低概括性.
相关概念视频
03:31End-To-End Deep Neural Network for Salient Object Detection in Complex Environments
1.0K
The present protocol describes a novel end-to-end salient object detection algorithm. It leverages deep neural networks to enhance the precision of salient object detection within intricate environmental...
1.0K
05:51Smartphone Fundus Photography
40.1K
Fundus photography normally requires specialized fundus cameras that are not always available in all clinical settings. Here, a simple method to record ocular fundus images using a smartphone camera and a conventional high-plus handheld indirect ophthalmoscopy lens is...
40.1K
04:55In Vivo Imaging of Retinal Microglia in a Mouse Model of Glaucoma
613
Source: Bosco, A., et al. In Vivo Dynamics of Retinal Microglial Activation During Neurodegeneration: Confocal Ophthalmoscopic Imaging and Cell Morphometry in Mouse Glaucoma. J. Vis. Exp. (2015) This video demonstrates in vivo imaging of retinal microglia in a mouse model of glaucoma. A transgenic mouse with fluorescent protein-expressing retinal microglia is positioned under an ophthalmoscope. Infrared illumination is used to locate the optic nerve head (ONH), and the focal...
613
07:22Quantitative Fundus Autofluorescence for the Evaluation of Retinal Diseases
11.9K
The retinal pigment epithelium (RPE) supports the sensory retina through recycling visual cycle byproducts, which accumulate as lipofuscin. These products are autofluorescent and can be qualitatively imaged in vivo. Here, we describe a method to quantitatively image RPE lipofuscin using confocal scanning laser ophthalmoscopy.
11.9K
12:06Analyzing Mitochondrial Morphology Through Simulation Supervised Learning
4.7K
This article explains how to use simulation-supervised machine learning for analyzing mitochondria morphology in fluorescence microscopy images of fixed cells.
4.7K
03:33Visualization of Neural and Vascular Networks in a Chicken Embryo
471
Source: Delalande, J., et.al. Dual Labeling of Neural Crest Cells and Blood Vessels Within Chicken Embryos Using ChickGFP Neural Tube Grafting and Carbocyanine Dye DiI Injection. J. Vis. Exp. (2015)This video demonstrates the transplantation of a GFP-labeled donor neural tube from a stage-matched transgenic chicken embryo into a recipient embryo at the level of somites one to seven, followed by vascular labeling using a lipophilic fluorescent dye. The combined approach allows for direct...
471