研究目的
To develop a method for the automated detection of new vessels from retinal images, which is crucial for the early detection of proliferative diabetic retinopathy (PDR), the advanced stage of diabetic retinopathy that carries a high risk of severe visual impairment.
研究成果
The described automated system is capable of detecting the presence of new vessels with high sensitivity and specificity, demonstrating its potential for clinical application in the early detection of proliferative diabetic retinopathy. The integration of ensemble classification and dual classification approaches has improved performance over previous methods.
研究不足
The system puts no emphasis on correctly detecting all new vessels, instead identifying any part of any new vessel region in the image is sufficient for the image to achieve a new vessel label. This approach may not delineate new vessel regions completely but ensures higher specificity. The lack of a standard dataset for testing makes true comparisons with other methods difficult.
1：Experimental Design and Method Selection:
The method is based on a dual classification approach where two vessel segmentation approaches are applied to create two separate binary vessel maps. Local morphology, gradient, and intensity features are measured to produce two separate 21-D feature vectors. Independent classification is performed for each feature vector using an ensemble system of bagged decision trees, and the outcomes are combined for a final decision.
2：Sample Selection and Data Sources:
Evaluation was performed using images from the publicly available MESSIDOR retinal image database and the St Thomas’ hospital ophthalmology department, totaling 60 images (20 with confirmed new vessels and 40 without).
3：List of Experimental Equipment and Materials:
Images were acquired from a color video 3CCD camera on a Topcon TRC NW6 fundus camera with a 45 degree field of view (FOV) and an image resolution of 2240 × 1488 pixels, and a Nikon D80 digital SLR camera on a Topcon TRC NW6 fundus camera with a 45 degree FOV and an image resolution of 2896 × 1944 pixels.
4：Experimental Procedures and Operational Workflow:
Images were spatially normalized, pre-processed, and then split into two pathways for vessel segmentation. Feature extraction was performed on each binary vessel map, followed by independent classification and combination of outcomes.
5：Data Analysis Methods:
Performance was assessed on a per image basis using sensitivity, specificity, and accuracy measures. The receiver operating characteristic (ROC) curve was used to visualize performance, and the area under the curve (AUC) was extracted as a performance measure.
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
