1. NEW APPLICATIONS OF WATERSHED SALIENCY
TECHNIQUE TO SOLVE OVER-SEGMENTATION
PROBLEM IN MEDICAL IMAGING
Muhannad Al-Hasan, Mark Fisher, Moe Razaz
School of Computing Sciences, University of East Anglia, Norwich, England
E-mail: m.al-hasan@uea.ac.uk, mhf@cmp.uea,ac.uk, mr@cmp.uea.ac.uk
Keywords: Watershed, Watershed segmentation, over-
segmentation, watershed algorithms, medical imaging.
Abstract
This paper aims to present some practical problems arising
from the watershed technique and address these by revisiting
the concept of watershed saliency based on a graph of the
mosaic image. Additionally, some new ideas and techniques
have been included to suppress over-segmentation. Firstly we
present a review of the watershed transform in the context of a
morphological segmentation. We then introduce the idea
behind a mosaic image, derived from an immersion process or
flooding simulation analogy. A weighted watershed algorithm
is introduced to overcome the problem of over-segmentation
in a standard watershed transform, and some new application
results are briefly presented and discussed.
1 Introduction
Image segmentation is a key problem in image processing and
is particularly important in medical imaging used for pre-
operative planning. Segmentation is the process of spatially
partitioning the pixels in any given image into homogeneous
classes with respect to some property such as intensity or
texture (texture refers to a pattern of closely placed elements
in such a manner that the pattern somehow repeats itself). A
good segmentation may be recognised from the
characteristics of its output components: each component
should be spatially cohesive as well as spatially accurate
while different components should be dissimilar [2]. In the
medical field, advances in imaging modalities such as CT and
MRI have enabled multidimensional visualisation of the
internal structures inside the human body. Visualisation of a
particular organ requires its extraction from the image using
segmentation techniques. There are many segmentation
techniques, the simplest of which is thresholding; however, in
many applications, choosing an accurate thresholding level is
usually difficult. The watershed is a popular segmentation
technique in the field of mathematical morphology [1]. In
order to understand the watershed, it is necessary to consider
the image as a surface, where high pixel values correspond to
peaks and low pixel values correspond to valleys. Just as with
actual watersheds, if a drop of water were to fall on any point
of the contour it would find its way to lower ground until it
reaches a local minimum. These local minima are referred
to as catchment basins, and all points that drain into the
same catchment basin are referred to as members of the
same watershed [3]. Several techniques have been used for
constructing watersheds, see for example [1,3].
2 Watershed Segmentation
Mathematical morphology provides a powerful approach
for segmentation through techniques based on the
watershed transform. The transform is usually applied to
gradient images in order to find the crest lines between the
different regional minima in the image. Although this
technique was developed in 1979, it is only recently that
the transform has become a popular tool for segmentation.
An analysis of various techniques used in watershed
segmentation has been presented in [5] by Hagyard and
Razaz. In its most general form, the watershed technique
can be applied to any image; however, in many
applications the method produces over- segmentation of the
objects in the image and their background. Marker images
are then used in conjunction with the watershed transform
to segment an image. Over-segmentation is particularly
problematic in the segmentation of medical image data into
anatomical structures, which is an important step for many
medical imaging applications that are currently under
development.
Watershed segmentation derives its name from the
manner in which the algorithm segments regions of an
image into “catchment basins” corresponding to local
minima. The regions surrounding these basins share
boundaries with one another as shown in Figure 1.
The output of the watershed algorithm is then a
hierarchy of basins, which can be viewed at different scale
levels. What remains is for the human to determine which
basins and at what scale levels represent the anatomical
structure or structures of interest.
The input to the watershed algorithm is critical to the
quality of its output image. The watershed algorithm
expects a “height image” as input, which is defined in this
context as an image where higher image intensities
correspond to logical boundaries between regions of
interest. The user needs to provide this input image to the
2. algorithm in order to highlight the important image features
for a given application. Thus, pre-processing the input image
to produce a “height image” is a critical step, often ignored by
some researchers in this field.
Figure 1: Catchment basins and watershed lines in watershed
segmentation.
3 Hierarchical Segmentation
In some cases, the markers selection and extraction are not so
easy. Some pictures may be very noisy and image processing
becomes more and more complex. The mosaic image has
been proposed as a possible marker-less solution to the over-
segmentation problem [4].
3.1 Suppression of Over-segmentation
Although the standard watershed algorithm produces an over-
segmented image, the contours in the image appear to be
correct. The main problem is how to choose the “right”
contours. The way we chose to tackle this problem is to by
using a mosaic image, which is achieved by post processing
the watershed algorithm proposed by Vincent and Soille [3].
3.2 The Mosaic Image
Consider a greyscale image f as in Figure 3(a), and its
morphological gradient image g(f) shown in Figure 3(b). Let
W(g) be the watershed transformation of g. Each catchment
basin produced by the transform is associated with a
minimum mi of the gradient g. A new function f /
, the mosaic
image of f, is defined by colouring each catchment basin
with its minimum pixel value. The mosaic image created in
this way is shown in Figure 3(c). A region adjacency graph
(RAG) as in Figure 2 is then derived from the mosaic image,
where the vertices of the directed graph correspond to the
catchment basins and the (weighted) edges represent the
saliency of watershed lines. Post-processing of this graph
can reduce over-segmentation.
Figure 2: Planar graph, adding vertices
corresponding to the primary catchment basins to provide
an intermediary connection between the neighbouring
vertices of the non planar graph.
3.3 Weighted Watershed Algorithm
A weighted watershed segmentation (WWS) algorithm
was developed which effectively overcomes the problem
of over-scgmentation. A morphological gradient image, i.e.
the difference between the dilated and eroded input image
is first generated. This is equivalent to finding the
difference between the largest and the smallest value of the
neighbours of each pixel. To this a flooding simulation
algorithm is then applied to separate the image into its
constituent catchment regions. A mosaic image is
subsequently generated by labelling each of the catchment
regions with the colour of its lowest point of catchment.
This tessellated image is used to generate weighted
watershed lines. The weight of each line (edge) is
proportional to the difference in colour between adjacent
regions. Thus the catchment regions associated with
homogeneous or background will not be deep and so will
have very weak watershed lines. The corresponding nodes
then are merged. But watershed lines separating objects
from background will have much stronger weights as they
separate catchment regions with a big colour difference.
This process results in watershed lines that are related to
the sharpness of the boundaries and hence results in
significant reduction of over-segmentation. Thresholding
the weighted watershed image will remove the irrelevant
segmentation lines, and will also even out the colouration
of the lines around a single object.
4 Experimental Results
We present some typical segmentation results using the
WWS algorithm. Figure 4a shows a dental X-ray image.
The image has already had some morphological operations
applied to it, specifically an opening to remove some black,
i.e. high intensity scratches. The bright white areas are
3. fillings in the patient’s teeth. These fillings have a variety of
shapes and can ‘blend’ into the areas on the image.
This image was first dilated and eroded and the results were
subtracted to produce a morphological gradient image (see
Figure 4b). The weighted watershed algorithm was applied
to this image to produce the results shown in Figures 4c and
4d. The mosaic image (Figure 4c) appears to be considerably
less complex than the original X-ray image, with the fillings
looking like one homogenous block.
Figure 3a Original image
Figure 3b Watershed segmented image
Figure 3c Mosaic image
Figure 3: MRI brain image
The weighted watershed segmented image as in Figure 4d
has clearly overcome the over-segmentation problem. In
this image the darkest watershed lines follow the
boundaries between catchment regions that have strongly
differing values in the mosaic image. Thresholding can
also be used to further enhance the segmented image if
needed.
Figure 4a Original dental X-ray image.
4. Figure 4b Morphological gradient image
Figure 4c Mosaic image of Figure 4a
References
[1] Beucher, S. 1994. Watershed, Hierarchical Segmentation
and Waterfall Algorithm. In: Serra,J.a.S.P. (ed).
Mathematical Morphology and Its Applications to Image
Processing. Kluwer Academic Publishers, pp. 69-76.
[2] Haralick, R. and Shapiro L. Survey: Image segmentation
techniques. 29. 1985. Computer Vision, Graphics, and Image
Processing.
Figure 4d Segmented dental image using Weighted
Watershed Algorithm
Figure 4: Weighted watershed segmentation of a dental
X-ray image
[3] Vincent, L. and Soille, P. 1991. Watersheds in Digital
Spaces - An Efficient Algorithm Based on Immersion
Simulations. Ieee Transactions on Pattern Analysis and
Machine Intelligence 13: 583-598.
[4] Najman,L. and Schmitt, M... 1996. Geodesic saliency
of watershed contours and hierarchical segmentation.
IEEE Transactions on Pattern Analysis and Machine
Intelligence 18: 1163-1173.
[5] Hagyard, D.M.P. and Razaz, M, 1996. “Analysis of
Watershed Algorithms for Greyscale Images", Proc. IEEE
I.C Image Processing, Vol. I.