Scholarlink

This paper gave insights about the deep learning methodologies to classification of retinal fundus images, thereby enhancing the accuracy and efficiency of eye disease detection. Some refinements were done in ODIR-5K dataset where distinct labels were assigned to each disease category to facilitate the classification. Data preprocessing techniques like Image preprocessing and data augmentation were implemented on dataset to standardize image quality. Among the evaluated pre-trained models, EfficientNet, DenseNet121, and MobileNet-V2 demonstrated superior performance, achieving 84.63% evaluation accuracy and 100% training accuracy. ResNet50 and VGG16 also provided the notable results, with evaluation accuracies of 83.44% and 81.99%, respectively. Optimal training parameters were identified through hyperparameter experimentation, while k-fold cross-validation was utilized to assess model generalization. The outcomes of this research have the potential to contribute to the development of clinical diagnostic tools for expeditious and accurate classification, ultimately leading to improved patient care.

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

Comparative Evaluation of CNN and Transformer Models for Image Classification: Insights from the ODIR 5K Dataset

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Anirudh Mukkamala,
  • Rishabh Chaudhary,
  • Vikram Singh Chauhan,
  • Zameer Fatima

摘要

This paper gave insights about the deep learning methodologies to classification of retinal fundus images, thereby enhancing the accuracy and efficiency of eye disease detection. Some refinements were done in ODIR-5K dataset where distinct labels were assigned to each disease category to facilitate the classification. Data preprocessing techniques like Image preprocessing and data augmentation were implemented on dataset to standardize image quality. Among the evaluated pre-trained models, EfficientNet, DenseNet121, and MobileNet-V2 demonstrated superior performance, achieving 84.63% evaluation accuracy and 100% training accuracy. ResNet50 and VGG16 also provided the notable results, with evaluation accuracies of 83.44% and 81.99%, respectively. Optimal training parameters were identified through hyperparameter experimentation, while k-fold cross-validation was utilized to assess model generalization. The outcomes of this research have the potential to contribute to the development of clinical diagnostic tools for expeditious and accurate classification, ultimately leading to improved patient care.