研究目的
To develop a multi-stage region-based convolutional neural network method for accurate detection and segmentation of the optic disc and fovea in retinal images, which is crucial for automated retinal disease detection.
研究成果
The proposed multi-stage faster-RCNN method effectively detects and segments the optic disc and fovea, with the RPI-based approach significantly improving fovea detection performance over standard methods, demonstrating its potential for automated retinal disease screening.
研究不足
The method may have limitations in cases where the fovea is not clearly visible or in images with severe abnormalities; Euclidean distance threshold of 400 pixels for undetected cases; reliance on the accuracy of OD detection for fovea detection.
1：Experimental Design and Method Selection:
A multi-stage approach using faster-RCNN for object detection and SVM for segmentation, with a novel RPI-based faster-RCNN for fovea detection.
2：Sample Selection and Data Sources:
Dataset from Indian Diabetic Retinopathy Image Dataset (IDRiD) with 467 fundus images; 413 for OD and fovea detection (dataset A), 54 for OD segmentation (dataset B). Images acquired using a Kowa VX-10 alpha digital fundus camera with 50-degree FOV, resolution 4288×2848 pixels.
3：List of Experimental Equipment and Materials:
Kowa VX-10 alpha digital fundus camera, computer systems for running neural networks (ResNet-50 base net), SVM implementation.
4：Experimental Procedures and Operational Workflow:
Preprocessing with histogram matching, OD detection using faster-RCNN with ResNet-50, OD segmentation using SVM based on pixel intensities, fovea detection using RPI-based faster-RCNN that incorporates relative position information from OD center.
5：Data Analysis Methods:
Evaluation using Euclidean distance for detection accuracy, Jaccard and dice indices for segmentation performance, five-fold cross-validation.
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
