Scholarlink

Visual impairment caused by several diseases can be reduced. Transition from manual methods to automation of diagnosis will enable rapid and efficient treatment in the early stages. To this end, we propose in our study a model (PfeTL) for the classification of three eye diseases: Diabetic Retinopathy (DR), Glaucoma (G), and Macular Degeneration (AMD), in addition to the normal category. Based on transfer learning and feature concatenation via the exploitation of EfficientNetB7 pre-trained models, CBAM attention mechanisms, and several convolutions, max-pooling, and fully connected layers. We will use 800 high-resolution images of fundus photography (CFP) from the public access dataset FIVES. An HP EliteBook 830 G8 Core i7 11th, 32 GB RAM is employed, with just the use of original data due to the very long time (more than 1 h 30 min) needed for every epoch in the case of data augmentation. Our proposed model achieves 80.00% in accuracy and a precision of 85.18% after 75 epochs.

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

Classification of Multiple Eye Diseases, Parallel Feature Extraction with Transfer Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Mohamed Chahid,
  • Abdelkarim Zemmouri,
  • Anass Barodi,
  • Mohsine Kartita,
  • Mohammed Benbrahim

摘要

Visual impairment caused by several diseases can be reduced. Transition from manual methods to automation of diagnosis will enable rapid and efficient treatment in the early stages. To this end, we propose in our study a model (PfeTL) for the classification of three eye diseases: Diabetic Retinopathy (DR), Glaucoma (G), and Macular Degeneration (AMD), in addition to the normal category. Based on transfer learning and feature concatenation via the exploitation of EfficientNetB7 pre-trained models, CBAM attention mechanisms, and several convolutions, max-pooling, and fully connected layers. We will use 800 high-resolution images of fundus photography (CFP) from the public access dataset FIVES. An HP EliteBook 830 G8 Core i7 11th, 32 GB RAM is employed, with just the use of original data due to the very long time (more than 1 h 30 min) needed for every epoch in the case of data augmentation. Our proposed model achieves 80.00% in accuracy and a precision of 85.18% after 75 epochs.