Scholarlink

Glaucoma is a leading cause of irreversible blindness worldwide. Early detection is crucial to prevent vision loss, but traditional diagnostic methods are subjective and prone to variability. This study compares the performance of three pre-trained convolutional neural network models: ResNet50, VGG16, and InceptionV3 in classifying glaucoma from images of nonglaucoma fundus using the ACRIMA dataset. The models were evaluated based on accuracy, recall, precision, F1 score, and AUC-ROC metrics. InceptionV3 demonstrated superior performance, achieving perfect classification (100% accuracy and AUC-ROC), while both ResNet50 and VGG16 reached excellent results (97.86% accuracy) with slightly different error patterns. Computational efficiency analysis revealed significant differences, with ResNet50 requiring 74 min for training, InceptionV3 95 min, and VGG16 228 min. These findings indicate that while InceptionV3 offers optimal diagnostic accuracy, ResNet50 provides an excellent balance between performance and computational efficiency, making it potentially more suitable for resource-constrained clinical environments.

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

Deep Learning Approaches for Glaucoma Detection: A Comparative Study of CNN Models on Retinal Fundus Images

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Eugenia Arrieta Rodríguez,
  • José Araque-Gallardo,
  • Oscar Luis Teheran Forero,
  • Emiro De-La-Hoz-Franco,
  • José Escorcia-Gutierrez

摘要

Glaucoma is a leading cause of irreversible blindness worldwide. Early detection is crucial to prevent vision loss, but traditional diagnostic methods are subjective and prone to variability. This study compares the performance of three pre-trained convolutional neural network models: ResNet50, VGG16, and InceptionV3 in classifying glaucoma from images of nonglaucoma fundus using the ACRIMA dataset. The models were evaluated based on accuracy, recall, precision, F1 score, and AUC-ROC metrics. InceptionV3 demonstrated superior performance, achieving perfect classification (100% accuracy and AUC-ROC), while both ResNet50 and VGG16 reached excellent results (97.86% accuracy) with slightly different error patterns. Computational efficiency analysis revealed significant differences, with ResNet50 requiring 74 min for training, InceptionV3 95 min, and VGG16 228 min. These findings indicate that while InceptionV3 offers optimal diagnostic accuracy, ResNet50 provides an excellent balance between performance and computational efficiency, making it potentially more suitable for resource-constrained clinical environments.