研究目的
To propose an effective recombined residual convolutional neural network (CNN) for recognizing AMD, DME, and normal SD-OCT images with less computation.
研究成果
The recombined residual CNN shows better performance than the original Residual CNN and related works in recognizing macular disorders when kernel sizes are relatively small, achieving the highest overall accuracy up to 90%. Detailed mid-level features increase recognition accuracy, but for DME, high-level features with larger receptive fields are also helpful. Future work will combine high-level features with larger receptive fields to the recombined residual CNN.
研究不足
The study is limited by the small scale of medical SD-OCT datasets compared to public datasets like ImageNet and COCO, which may lead to poor generalization or over-fitting. The performance advantage of the recombined residual CNN reverses when kernel size is fixed to 7x7.
1：Experimental Design and Method Selection:
The study adopts an 18-layer Residual CNN as the basis, recombining it by removing some convolutional layers from groups extracting low-level or high-level features and adding them to the middle part to enhance mid-level feature extraction. Kernel sizes are adjusted to change visual receptive fields during convolution calculations.
2：Sample Selection and Data Sources:
The SD-OCT dataset used is publicly available, obtained from Duke University, Harvard University, and the University of Michigan, consisting of images from 45 subjects (15 normal, 15 AMD, 15 DME).
3：List of Experimental Equipment and Materials:
Intel Core i7-6700HQ CPU with 8GB RAM, Caffe open-source library for implementation.
4：Experimental Procedures and Operational Workflow:
Images are denoised using BM3D filtering before being sent to the CNN. The network is pre-trained on ImageNet dataset. Training involves a learning rate of 0.01, divided by 10 every 1500 iterations, using stochastic gradient descent for optimization.
5：01, divided by 10 every 1500 iterations, using stochastic gradient descent for optimization.
Data Analysis Methods:
5. Data Analysis Methods: Performance is evaluated using 5-fold cross-validation, with metrics including accuracy, precision, recall, f1-score, and average time consumption per image.
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
