Scholarlink

<p>This research analyzes the effectiveness of various deep neural network (DNN) architectures for optical illness detection using the Ocular Disease Intelligent Recognition (ODIR) dataset. Preprocessing techniques, including bilateral filtering, unsharp masking, and region of interest (ROI) selection, were applied to enhance image quality and focus on critical features. The performance of multiple models—ResNet50, InceptionV3, MobileNet, DenseNet121, NASNetMobile, Xception, and VGG19—was evaluated across different artificial neural network (ANN) architectures such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and DNNs. Our results show that ResNet50, leveraging its deep residual learning capabilities in CNN and LSTM setups, achieved an accuracy of up to 99.74%. InceptionV3 and Xception demonstrated strong performance in CNN-based frameworks due to their superior feature extraction techniques, achieving accuracies of 93.70% and 85.38%, respectively. MobileNet maintained a balance between computational efficiency and accuracy, particularly in CNN and DNN frameworks, reaching an accuracy of 96.40%. DenseNet121’s dense connectivity yielded strong results in both CNN and LSTM structures, with an accuracy of 98.88%, while NASNetMobile, optimized via reinforcement learning, performed well in DNNs, attaining up to 97.89% accuracy. VGG19 consistently produced reliable results across multiple neural network topologies. A comparison of the ADAM and RMSprop optimization algorithms indicated that ADAM generally outperformed RMSprop in precision, recall, F1-score, and accuracy, with average values of 95.79%, 94.21%, 95.02%, and 95.36%, respectively. RMSprop exhibited slightly lower mean values of 95.19%, 94.24%, 96.01%, and 94.84%. Detailed statistical analysis confirmed the stability and robustness of the ADAM optimizer, with reduced variability and skewness. This study highlights the significance of model architecture and optimizer selection in achieving high performance in ocular disease detection and provides valuable insights into the application of deep learning techniques for medical image analysis.</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

A Comprehensive Study on Deep Learning Models for the Detection of Diabetic Retinopathy using Pathological Images

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Davinder Paul Singh,
  • Tathagat Banerjee,
  • Shubham Mahajan,
  • K. Ramesh Chandra,
  • Raju Kumar,
  • Shanta Phani,
  • Pawandeep Kour,
  • Yogendra Narayan,
  • Gurpreet Kour,
  • Loveleena Mukhija,
  • Debabrata Swain,
  • Malvika Ashok

摘要

This research analyzes the effectiveness of various deep neural network (DNN) architectures for optical illness detection using the Ocular Disease Intelligent Recognition (ODIR) dataset. Preprocessing techniques, including bilateral filtering, unsharp masking, and region of interest (ROI) selection, were applied to enhance image quality and focus on critical features. The performance of multiple models—ResNet50, InceptionV3, MobileNet, DenseNet121, NASNetMobile, Xception, and VGG19—was evaluated across different artificial neural network (ANN) architectures such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and DNNs. Our results show that ResNet50, leveraging its deep residual learning capabilities in CNN and LSTM setups, achieved an accuracy of up to 99.74%. InceptionV3 and Xception demonstrated strong performance in CNN-based frameworks due to their superior feature extraction techniques, achieving accuracies of 93.70% and 85.38%, respectively. MobileNet maintained a balance between computational efficiency and accuracy, particularly in CNN and DNN frameworks, reaching an accuracy of 96.40%. DenseNet121’s dense connectivity yielded strong results in both CNN and LSTM structures, with an accuracy of 98.88%, while NASNetMobile, optimized via reinforcement learning, performed well in DNNs, attaining up to 97.89% accuracy. VGG19 consistently produced reliable results across multiple neural network topologies. A comparison of the ADAM and RMSprop optimization algorithms indicated that ADAM generally outperformed RMSprop in precision, recall, F1-score, and accuracy, with average values of 95.79%, 94.21%, 95.02%, and 95.36%, respectively. RMSprop exhibited slightly lower mean values of 95.19%, 94.24%, 96.01%, and 94.84%. Detailed statistical analysis confirmed the stability and robustness of the ADAM optimizer, with reduced variability and skewness. This study highlights the significance of model architecture and optimizer selection in achieving high performance in ocular disease detection and provides valuable insights into the application of deep learning techniques for medical image analysis.