Comparative study of retinal vessel segmentation methods on a new publicly available database | Niemeijer et al.

For ISMI, focus on method described in 2.3.5.

Golden standard of manually segmented retinal vessels.

Automatic retinal vessel segmentation is useful because:

  • Useful for monitoring the vessels over time
  • Useful for excluding irrelevant sections of the image for easier analysis
  • Position, shape and size of vasculature can be used to find the optic disk and the fovea

Compared Methods

  1. Matched Filter Approach
    • Cross-sections can be approximated by a 2D Gaussian filter
    • Convolve image with 15 degree rotations of the filter, keep maximum response
    • Threshold result for binary segmentation
  2. Scale-space analysis and region growing
    • Features, both at different scales: gradient magnitude of intensity; ridge strength
    • ridge strength: absolute largest eigenvalue of Hessian (matrix of second order derivatives)
    • normalize by the used scale and retain local maxima
    • histogram of features as input to region growing algorithm
  3. Mathematical Morphology and Curvature Estimation
    • Math heavy: based on "mathematical morphology"; couldn't be bothered.
  4. Verification-Based Local Adaptive Thresholding
    • Global threshold is not suitable because the background of retinal images have strong gradients
    • Instead probe image with several thresholds
    • Do binary segmentation for each threshold
    • Classify the segmented objects, retain vessel-like ones
    • "All the retained binary objects can be combined to form a binary vessel tree segmentation" (5))
  5. Pixel Classification Approach
    • The only supervised method
    • Extract feature vector for each pixel
    • Features only extracted from the green plane
    • Gaussian filter and 1-st and 2-nd order derivatives with s=1,2,4,8,16
    • With derivatives in both directions, 5x6+1=31 features
    • Feature centered to zero mean and unit variance
    • kNN classifier with k=30
    • kNN allows soft classification, i.e. output a probability map with a probability for each pixel that is is a vessel pixel
    • Binary segmentation achieved with a threshold on the probability p

Evaluation showed that automatic segmentation of smaller vessels is especially problematic, partly due to grey smudgy edges and JPEG compression.