研究目的
To propose RetoNet, a deep learning architecture for automated detection of retinal ailments from fundus images, aiming to provide a cost-effective and convenient diagnostic tool for age-related macular degeneration and diabetic retinopathy.
研究成果
RetoNet achieved high accuracy in detecting retinal ailments, demonstrating the potential of deep learning for automated healthcare diagnostics. Future work could involve larger datasets and distributed deep learning for scalability.
研究不足
The dataset used is not large, which may limit performance; public access to medical image datasets is restricted due to confidentiality; transfer learning with VGGNet did not perform optimally due to dataset dissimilarity.
1：Experimental Design and Method Selection:
The study designed and trained a convolutional neural network (CNN) architecture called RetoNet from scratch, and compared it with a transfer learning model based on VGGNet. The methodology involved data preprocessing, augmentation, and optimization using RMSProp.
2：Sample Selection and Data Sources:
The ARIA dataset was used, consisting of fundus images from 61 healthy subjects, 59 with diabetic retinopathy, and 23 with AMD. Images were captured using a Zeiss FF450+ fundus camera.
3：List of Experimental Equipment and Materials:
A Zeiss FF450+ fundus camera for image capture; computational resources on Microsoft Azure cloud; software tools including Keras with TensorFlow backend.
4：Experimental Procedures and Operational Workflow:
Images were converted from TIFF to JPEG, normalized, and augmented by flipping and rescaling. The dataset was split 80:20 for training and testing. RetoNet was trained for 150 epochs with hyperparameter tuning.
5：Data Analysis Methods:
Performance was evaluated based on accuracy, loss plots, and comparison with existing methods using metrics like accuracy percentages.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
