Abstract

 

Purpose

 

To describe and test a novel, robust method for automatic segmentation of vessels in complicated color retinal images with confounding vessel-like objects.

 

Methods

 

This method classifies pixels of 2D fundus images as vessel or non-vessel pixels. Twenty 35 degree field of view photographic retinal images from the STARE database collected at the UC San Diego, manually annotated by an expert (MHG), were used as a training and test set for the classifier. An example STARE image and its annotated ground truth vessels are shown in Fig1.a. Intensities of pixels in the square neighborhood of pixels were used to build raw feature vectors. The square neighborhood of size 9x9 pixels was used to build an 81-dimensional raw feature set for every pixel. The square neighborhood for pixel p is highlighted with yellow window in Fig1.a. The 81-dimensional feature set was reduced to 7 dimensions through a two-stage dimensional reduction process as shown in Fig1.b. In the first stage, principal components analysis (PCA) was applied to reduce dimensions of feature set and then linear discriminant analysis (LDA) was employed in the second stage to map vessel and non-vessel pixels into new spaces that are more discriminant. The classifier is constructed based on ensembles of n decision trees as shown in Fig1.c. To provide independent training and test image sets, leave-one-out methodology was used to train classifiers on 19 images and to test on one image with the process repeated 20 times to test all images. In our experiments n=60 and final classification is obtained using majority vote. The classified image is further refined by applying post processing to remove outliers.

 

Results

 

The proposed classifier is 88.98% sensitive and 91.05% specific for vessel classification in complicated images before post processing and it is improved to 90.10% and 90.13% respectively after post processing with the area under the curve (ROC) of 0.9006. The example of STARE image, ground truth and segmented vessels by our method is shown in Fig.2.

 

Conclusions

 

The proposed method classifies pixels of 2D fundus images as vessel or non-vessel. It uses patch based PCA and LDA to extract features and classifies them using ensembles of decision trees. The results show vessels are segmented effectively in normal and complicated images.