Scholarlink

The human eye is highly sensitive and vulnerable to various diseases that need timely attention; otherwise, they may lead to vision impairment. According to WHO (World Health Organization), most of these cases could be avoided through regular examinations. Therefore, early detection of eye disease has become a necessity. In this manuscript we propose a novel approach for prediction and classification of ocular diseases like age-related macular degeneration, cataract, diabetic retinopathy, glaucoma, hypertensive retinopathy, myopia, and normal using fundus images collected from multiple public sources. Our study comprises the study of multiple contemporary deep learning models, including EfficientNet-B4, ConvNeXt, DenseNet-201, and ViT-16B. Our proposed solution involves the implementation of the ConvNeXt model on our dataset, which achieves an accuracy of 94.14%. Our purpose behind this study is not only to develop an effective classification model but also to ensure its visual – understanding how and why the model is making certain decisions. To achieve this, we implemented the Gradient-Weighted Class Activation Mapping (Grad-CAM) to provide an explainable AI (XAI) based solution to help researchers and practitioners in the field.

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

OcuXPlain: An Explainable AI Approach for Multi-class Ocular Disease Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Manan Parekh,
  • Anmol Aafre,
  • Parth Patel,
  • Vaidehi Lad,
  • Nirali Nanavati

摘要

The human eye is highly sensitive and vulnerable to various diseases that need timely attention; otherwise, they may lead to vision impairment. According to WHO (World Health Organization), most of these cases could be avoided through regular examinations. Therefore, early detection of eye disease has become a necessity. In this manuscript we propose a novel approach for prediction and classification of ocular diseases like age-related macular degeneration, cataract, diabetic retinopathy, glaucoma, hypertensive retinopathy, myopia, and normal using fundus images collected from multiple public sources. Our study comprises the study of multiple contemporary deep learning models, including EfficientNet-B4, ConvNeXt, DenseNet-201, and ViT-16B. Our proposed solution involves the implementation of the ConvNeXt model on our dataset, which achieves an accuracy of 94.14%. Our purpose behind this study is not only to develop an effective classification model but also to ensure its visual – understanding how and why the model is making certain decisions. To achieve this, we implemented the Gradient-Weighted Class Activation Mapping (Grad-CAM) to provide an explainable AI (XAI) based solution to help researchers and practitioners in the field.