Retinal Vessel Centerline Extraction Using Multiscale Matched Filters, Confidence
and Edge Measures
This page gives a high level overview of our advances in the retinal extraction work.
For more details, please refer to our article
in IEEE Transactions on Medical Imaging.
Motivated by the goals of improving detection of low-contrast and
narrow vessels and eliminating false detections at non-vascular
structures, a new technique is presented for extracting vessels in
retinal images. The core of the technique is a new likelihood
ratio test that combines matched-filter responses, confidence
measures and vessel boundary measures. Matched filter responses
are derived in scale-space to extract vessels of widely varying
widths. A vessel confidence measure is defined as a projection of
a vector formed from a normalized pixel neighborhood onto a
normalized ideal vessel profile. Vessel boundary measures and
associated confidences are computed at potential vessel
boundaries. Combined, these responses form a 6-dimensional
measurement vector at each pixel. A training technique is used to
develop a mapping of this vector to a likelihood ratio that
measures the ``vesselness'' at each pixel. Results comparing this
vesselness measure to matched filters alone and to measures based
on the Hessian of intensities show substantial improvements both
qualitatively and quantitatively. The Hessian can be used in place
of the matched filter to obtain similar but less-substantial
improvements or to steer the matched filter by preselecting kernel
orientations. Finally, the new vesselness likelihood ratio is
embedded into a vessel tracing framework, resulting in an
efficient and effective vessel centerline extraction algorithm.
Reliable vessel extraction is a prerequisite for subsequent retinal
image analysis and processing because vessels are the predominant and
most stable structures appearing in the images. Several challenges of vessel
extraction in retinal images are
illustrated by the images shown in Figures 1 and 2. These challenges may be outlined
- There is a wide range of vessel widths -- from less than a
pixel to 12 pixels wide in the example shown.
- Vessels may be low contrast. The central intensity of some
vessels differ from the background by as little as 4 grey levels,
yet the background noise standard deviation is 2.3 grey levels.
Narrow vessels often have the lowest contrast.
- A variety of structures appears in the images, including the
retina boundary, the optic disc, and pathologies. The latter are
particularly challenging for automatic vessel extraction because they
may appear as a series of bright spots, sometimes with narrow,
darker gaps in between.
- Wider vessels sometimes have a bright strip running down the
center (the ``central reflex''), causing a complicated intensity
cross-section. Locally, this may be hard to distinguish from two
Our focus in this work is on techniques needed to solve the first
three problems -- detecting low-contrast vessels and narrow
vessels, while avoiding false responses near pathologies and other
Illustration of the challenges of retinal vessel extraction. Arrows
drawn on the image [in yellow / with dashed lines] indicate
pathologies, the boundary of the optic disc and the boundary of the
retina, all of which tend to cause false positives in vessel detection.
Arrows drawn [in white / with solid lines] highlight narrow or
low-contrast vessels which are difficult to detect.
The limitations of some existing methods are illustrated in
Figure 2. All these ``vesselness measures''
produce substantial responses to non-vascular structures such as
the optic disc and the pathologies. All measures produce stronger
responses at the boundary of the retina, near the optic disc, and
along central pathological structures than for the thin and
The main contribution of this paper is the development of an
enhanced vesselness measure that addresses the issues illustrated
in Figure 1. Motivated by the apparent
effectiveness of the matched filter in highlighting low-contrast
and narrow vessels and by recent success in using matched filters
for retina vessel segmentation , we introduce a
multi-scale matched filter for vessels, using an appropriate
normalizing multiplier to allow the combination of responses
across scales. We then augment the matched-filter responses with a
new vessel ``confidence'' measure, analogous to the edge-based
measure presented in . It can be viewed as normalized
cross-correlation or cosine distnace between model and sample vectors and
determines how closely
an image region follows an ideal vessel profile. Importantly,
unlike the matched filter, this measure is independent of
amplitude. To these vessel response and confidence values we add
edge detection filter responses and confidences taken from the
boundary of the purported vessel. This produces a six
degree-of-freedom measurement vector at each pixel. Then, we use
a training technique to develop a mapping from this vector to a
single likelihood ratio that serves as the final ``vesselness''
measure. This gives a measure at each pixel which may be used
either for segmentation of vessel pixels or for identifying the
centerline vessel pixels and vessel widths when used in
combination with non-maximum suppression. We focus on the latter
because the measures are designed to have maximum response along
the centerline of the vessel, and because this provides a more
compact, geometric description of the vessels than segmentation
We show that the new Likelihood Ratio Vesselness (LRV) measure
outperforms the multiscale matched-filter and existing
Hessian-based measures using quantitative analysis and using
visual inspection of results on retinal images. We also show how
the multiscale Hessian may be used to steer the matched filter by
selecting its orientation at each pixel and scale. This way, the
matched filter is applied only once at each pixel and scale,
eliminating much of the excess computation. Alternatively, the
Hessian may be used in place of the matched-filter in the LRV
measure, producing substantial improvements in Hessian-based
vesselness measures. The advantage of this is a lower overall
computational cost than the matched-filter-based measure in
exchange for a slight decrease in effectiveness.
In our evaluation, we first quantify the improvement gained by
using vessel confidences and edge measures together with the
matched filter. We then compare detection results using LRV and
Frangi's measure on all retinal vessels in an image. We then
show qualitative comparison of the
LRV measure and the vessel matched filter.
Our quantitative analysis uses overall results
combined across all 20 images in the STARE data set
(Figure 3). The use of the
matched filter and its associated confidence -- a two-component
measurement vector at each pixel -- in forming the likelihood
ratio dramatically improves the performance of the matched filter,
while adding the edge responses and confidences needed to form the
complete 6-component vector provides significant further
improvements. Both the Hessian-based (Frangi)
and the matched-filter-based LRV measures substantially outperform
the original measures, but the matched-filter LRV is clearly
superior. Finally, the seemingly-strange non-monotonic shape of the
matched-filter and Frangi measures alone is easily explained. The
highest responses for these measures occur at the retina boundary,
the optic disc boundary, and the boundary of pathologies (though
perhaps slightly offset from the true boundary, as discussed
above), because they are of much higher-contrast than even the
more distinct vessels. Hence, for very high thresholds, these are
the only responses that survive thresholding.
(a) (1-Precision)-Recall curves
showing the effectiveness of the LRV measure.
Comparisons are made between the full measure (LRV), the measure
with only vessel confidences, and the multiscale matched filter
alone. (1-Precision)-Recall curves are more suitable for
comparison because of large number of negatives (non-vessel
pixels) in the ground truth than ROC curves. Notice that the
matched filter with vessel confidences is as powerful as the
full LRV measure until about 50% of all traces are detected.
Plot (b) compares the LRV measure (with the matched
filter) to the LRV using Frangi's Hessian-based
measure and to the matched filter and Frangi's measure alone.
Qualitative Results - Pathologies
In order to give a qualitative feel for the results, several
``chips'' from the most challenging images are shown in
Figure 4. These include a variety of pathologies
and thin vessels that are so subtle they sometimes completely fade
into the noise.
Results on 5 difficult image chips showing a variety of
challenges, including pathologies, other non-vascular
structures, and low-contrast, narrow vessels. The figure shows
the source images in the 1st column, the vessel
matched-filter response images after non-maximum suppression in the
2nd column, the Likelihood Ratio Vesselness (LRV) after
non-maximum suppression in the 3rd column and the
pixels that remain after thresholding the LRV measure at
= 2.0 in the 4th column.
Starting from a set of oriented seed locations, exploratory tracing algorithms
recursively trace the vascular structure by stepping along the vessel
tangent direction and locally detecting the next point on the
vessel. Usually, the ``centerline'' location, vessel width, and
tangent direction are estimated simultaneously. Tracing stops for
a given seed point when it intersects a vessel already traced or
when the vessel is not reliably detected. Our approach follows this strategy as well, with a small but
important variation. The natural approach would be to use the 6-D
measurement vector of matched-filter, confidence and edge
responses, computed over a small neighborhood in location,
orientation and scale to compute the likelihood ratio and then
apply a threshold to decide the presence of a vessel. We have
found, however, that this can lead to missed traces for the most
subtle vessels, prematurely halting tracing. Instead we apply only
the vessel matched filter during tracing. This is very sensitive
and detects both subtle vessels as well as results in many traces
at non-vessel locations. We then
compute the 6-D measurement and apply the likelihood ratio
subsequently to these ``over-traced'' results. In effect we are
using the vessel matched filter in tracing to generate candidate
locations at which to evaluate the LRV measure. This is less
expensive than starting from large number of seed points and
applying LRV measure during tracing.
Source image (1st column), tracing results using
parallel edge algorithm  (2nd column),
Likelihood Ratio Vesselness (LRV) based tracing (3rd
column) and the LRV after non-maximum suppression and
thresholding below 2.0 (4th column). True positives
are in blue, false negatives in red, and false positives in
green. The new tracing algorithm successfully handles difficult
vessel branching and crossover points (two overlapping vessels
in the 3rd and 4th example), more accurately determines when to
stop tracing in unclear regions (thin vessels near pathologies
in the 1st and 3rd example), correctly ignores pathologies (1st
and 4th example) and finds many of the thinnest vessels which
usually have very low contrast (5th example). Notice in a number
of cases false positives are marked alongside false negatives
(green line along red line) because of localization differences
with the ground truth. The tracing results are comparable to
the LRV measure evaluated at every point. Vessels with fewer
than 4 connected pixels were removed.
One final example that reinforces the effectiveness of LRV tracing
is shown in Figure 6 where both tracing algorithms
are applied to an image from a slit-lamp biomicroscope. The
likelihoods are powerful enough to highlight vessels but ignore
most of the background noise. This noise causes detection of
spurious traces by the parallel-edge based algorithm. The new
tracing algorithm correctly finds vessels that are missed by the
parallel-edge based tracing.
Example of a slit lamp image (top left), LRV measure
computed image wide (top right), parallel-edge based
tracing (bottom left), and LRV-based tracing
(bottom right). The LRV measure is powerful enough to
highlight vessels but ignore the background noise which is the
main reason for spurious traces of the parallel-edge based
algorithm. The new tracing algorithm correctly finds vessels
that are missed by the parallel-edge based tracing.
Each component of the Likelihood Ratio Vesselness (LRV) measure
is designed to help address the
problem of detecting narrow and low-contrast vessels, while
avoiding responses to other retinal image structures. In
particular, the elongated template of the multiscale matched
filter tends to preserve vessels that are only a pixel wide (and
typically low-contrast), whereas isotropic measures such as the
Hessian tend to substantially blur these structures. The edge
responses are useful in distinguishing between offset edges near
pathologies and true vessels. The confidence measures emphasize
the shape of the intensity surface rather than the magnitude of
the responses, enhancing the ability of the LRV to detect
low-contrast vessels. The 6-component measurement vectors are
mapped to the final LRV measure using pdfs learned from training
data. Significantly, after designing multiscale filtering and
training procedures there are no run time tuning parameters in the
computation of this measure. The computation is made more
efficient by using the Hessian to select matched-filter
orientations and by embedding the LRV measure in a tracing
framework. The latter substantially outperforms our existing
retinal vessel tracing algorithm. Both quantitative and
qualitative analysis on challenging retinal images have shown the
effectiveness of the LRV measure. The new measure may be used in
place of the Hessian and the matched filter in existing vessel
detection and segmentation algorithms [2,7]. Based on the experimental evaluations reported here,
this should lead to substantially-improved results. The fact that
the Hessian may be used to steer the application of the
matched-filter and confidence measures makes the computation
Publications and Further Reading
A. Can, H. Shen, J. N. Turner, H. L. Tanenbaum, and B. Roysam.
Rapid automated tracing and feature extraction from live
high-resolution retinal fundus images using direct exploratory algorithms.
IEEE Trans. Inform. Technol. Biomed., 3(2):125-138, 1999.
A. Frangi, W. J. Niessen, K. L. Vincken, and M. A. Viergever.
Multiscale vessel enhancement filtering.
In Proc. 1st MICCAI, pages 130-137, 1998.
A. Hoover, V. Kouznetsova, and M. Goldbaum.
Locating blood vessels in retinal images by piecewise threshold
probing of a matched filter response.
IEEE Trans. Med. Imag., 19(3):203-210, 2000.
Edge detection and ridge detection with automatic scale selection.
Int. J. Comp. Vis., 30:117-156, November 1998.
V. Mahadevan, H. Narasimha-Iyer, B. Roysam, and H.L. Tanenbaum.
Robust model-based vasculature detection in noisy biomedical images.
IEEE Trans. Inform. Technol. Biomed., 8(3):360-376, 2004.
Peter Meer and Bogdan Georgescu.
Edge detection with embedded confidence.
IEEE Trans. Pattern Anal. Machine Intell., 23(12):1351-1365, December 2001.
J.J. Staal, M.D. Abramoff, M. Niemeijer, M.A. Viergever, and B. van Ginneken.
Ridge based vessel segmentation in color images of the retina.
IEEE Trans. Med. Imag., 23(4):501-509, Apr 2004.