Scholarlink

<p>Diabetic retinopathy is a severe eye condition that can lead to vision loss at severe stages necessitating the early detection. Automating the detection reduces the labor and facilitates timely intervention. Deep learning models have advanced detection capabilities but often suffer from overfitting due to limited or imbalanced labeled fundus data. Most of the existing models still depend on conventional augmentation techniques which fail to provide a controlled realistic fundus image generation. To address these limitations, this paper presents Optic-GAN, a strategic, memory- and time-efficient generative model for dataset expansion. Optic-GAN balances the MESSIDOR I and II datasets, achieving a low Fréchet inception distance score and improving model accuracy from 89.56 to 96.06%. Optic-GAN stands out by being independent of the dataset structures, unlike existing models, enabling broader, more effective application across diverse contexts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optic-GAN: a generalized data augmentation model to enhance the diabetic retinopathy detection

  • 期刊论文

  • 发表日期：

    2025年2月13日

查看全文
  • Preeti Kapoor,
  • Shaveta Arora

摘要

Diabetic retinopathy is a severe eye condition that can lead to vision loss at severe stages necessitating the early detection. Automating the detection reduces the labor and facilitates timely intervention. Deep learning models have advanced detection capabilities but often suffer from overfitting due to limited or imbalanced labeled fundus data. Most of the existing models still depend on conventional augmentation techniques which fail to provide a controlled realistic fundus image generation. To address these limitations, this paper presents Optic-GAN, a strategic, memory- and time-efficient generative model for dataset expansion. Optic-GAN balances the MESSIDOR I and II datasets, achieving a low Fréchet inception distance score and improving model accuracy from 89.56 to 96.06%. Optic-GAN stands out by being independent of the dataset structures, unlike existing models, enabling broader, more effective application across diverse contexts.