Sie sind auf Seite 1von 5

Automatic Retinal Image Registration Using fully

Connected Vascular Tree


Janabai Parekar, Prasanna Porwal and Manesh Kokare
Center of Excellence in Signal and Image Processing
Shri Guru Gobind Singhji Institute of Engineering and Technology
Vishnupuri, Nanded (M.S.), India
{parekarjanabai14-15, porwalprasanna, mbkokare}@sggs.ac.in

Abstract—Retinal image registration plays very important role Feature-based method uses features of image as vascular
in medical imaging, as it combines the information of different bifurcation points, trifurcation points, whole vascular tree, and
field of view images and temporal images. This paper presents the other feature of image. Steward et al. [4] used vascular bifur-
novel method for retinal image registration based on bifurcation
structure matching. In proposed method, segmentation is based cation points and crossing points as landmark points. Yang et
on the multi-scale ridge detection, and the preprocessing steps al. [5] proposed Generalized Dual-Bootstrap Iterative closest
used to avoid false detection of vessels. Also, Dijkstra’s algorithm Point (GDB-ICP) method based on SIFT key points. Feature-
is used to get a fully connected vascular tree. A novel technique is based registration methods also fail to align low-quality retinal
proposed for removing the false correspondence. This method is images. Though there are different methods available for
invariant to translation, rotation and scaling since feature vector
contains normalized lengths and angles. The registered image is retinal image registration, it is still very challenging to find
visually good as well as experimental performance shows very pixel correspondence. These challenges are, firstly there is very
good results. small overlapping area between the images. Secondly, retinal
Index Terms—Retinal image registration, Vessel segmentation, images having bleeding or tumor tissue reduce the number
Bifurcation structure, Transformation model. of detectable bifurcation points or may inhibit their correct
segmentation. Thirdly, contrast and intensity distribution of
I. I NTRODUCTION images create problems for feature based and intensity based
methods. Finally, a difference of intensity between the vessel
Retinal imaging has significantly grown in the field of center line and vessel edges, which causes false detection of
ophthalmology in last few decades. Fundus photography is vessels.
the non-invasive method, which takes direct images of the To solve above challenges, we have proposed a novel
retina. Current research is focused on automatic early detection method using bifurcation structure which uses ordered statistic
of diabetic retinopathy using fundus photographs. Automated filter for processing, multi-scale ridge detection for vessel
retinal image registration is an important step in diabetic segmentation and Dijkstra’s algorithm for joining disconnected
retinopathy for treatment evaluation. An accurate registration vessels. This paper is organized as follows, Section II explains
helps to diagnose the progression of diseases. Hence, this the proposed method. In section III, experimental results are
paper focusses on retinal image registration. Image registration given, finally section IV concludes this paper.
is defined as aligning two or more images of the same scene
taken at a different time with different camera parameter. In II. P ROPOSED M ETHODOLOGY
this process, one image is taken as a reference image or fixed Fig. 1 shows proposed methodology. The retinal vessel has
image. The transformation parameter of another image (target different width and orientation, hence segmentation is not
image) with respect to the reference image is found and these possible using a single scale. So multi-scale ridge detector
two images are aligned. is used to segment vessels of different widths. In proposed
Broadly, retinal image registration belongs to two cate- method images are resized to 858×570 in order to reduce
gories. First category is intensity based and second is feature computational complexity.
based [1]. The intensity-based methods depend on gradient and
cross-correlation, intensity differences and shared information A. Preprocessing of Retinal Images
of images. Kim et al. [2] proposed a method which is insensi- Retinal blood vessel segmentation is challenging due to
tive to outliers. Matsopoulos et al. [3] used simulated annealing low contrast between blood vessels and retinal background
and genetic algorithms to improve the object function based hence, green channel is used for segmentation [6]. Vessels
on the intensity differences of the segmented retinal images. in retinal images are darker in nature, but the center line
Intensity-based method needs whole image information for of some vessels has high-intensity compared to their edges.
registration. If quality of image is a low or overlapping area Due to this, the output of segmentation contains edges of
in concerned images are lesser, then intensity based method vessels. To remove this effect ordered statistic filter is used.
might fail to register images. A minimum and maximum ordered filter is used successively.
Ixx + Iyy ± α
λ1 , λ2 = (2)
2
Where, Ixx , Iyy are second order derivative of image in x,
y directions respectively and
q
α = (Ixx − Iyy )2 + 4Ixy2 (3)
The response is pixel-wise maximum over different scales.
1) Non-maximum suppression: The output of the previous
step is a gray scale image having a maximum response to
vessels. To obtain vessel centerline, pixels around vessel center
line have to suppress. To decide it, an image is scanned along
the direction of the gradient, and if the pixel is not local
maxima then it is set as zero. After this step, a threshold is set
and the image is converted to a binary image. Morphological
thinning is applied on this image for getting single pixel width
vessels.
2) Noise Removal: Result of non-maximum suppression
Fig. 1. Flow chart of proposed method. contains false and disconnected vessels. These false vessels
are removed using following steps.
Step I: Determine the connected component.
Step II: Compute the area of each component.
Step III: Remove small object.
To find connected component eight neighborhood connectivity
is used. The component having an area less than 20 pixels are
removed, which is usually false vessels. Fig. 3(b) shows the
output of vessel segmentation after noise removal.
3) Connecting Vessel: Noise removed image still contains
some disconnected vessels. These vessels should connect such
that it must look natural. Dijkstra’s single source shortest path
algorithm is used to find the shortest path from source to every
(a) (b)
other vertex. Let G = (V, E) is a given graph, where V = vi
Fig. 2. Output of pre-processing : (a) Original image, (b) Image after is set of vertices, and E = wij is weight of arc from vi to
processing. vj . Time complexity of Dijkstra’s algorithm [8] having M
nodes is O(M 2 ) . We label each disconnected vessel as one
object. Suppose there are k numbers of disconnected vessels,
Minimum filtering blurs the vessel and fills the center line
so resultant image is formed by combining these isolated
of vessels, whereas maximum filtering is used to sharpen the
objects and background as:
vessel boundary. These preprocessing steps improve contrast
between vessels and background as well as remove edge Ω = Ω0 ∪ Ω1 ∪ ... ∪ Ωk ; ∀i 6= j, Ωi ∩ Ωj = Ø (4)
detection effect. Fig. 2 show the effect of preprocessing on
the retinal image. Here Ω0 is background and Ω1 ...Ωk is an isolated vessel in
foreground. We define the graph in which every image pixel
B. Vessel segmentation acts as vertex. If two vertices are four or eight neighborhoods
In order to proceed for registration, we need to segment of each other then they are connected by an arc directed from
retinal vessels first. As blood vessels are elongated structures vertex vi to vj . The distance between such vertices is given
of different diameters and orientation, vessel segmentation by weight of an arc from vi to vj as:
is mentioned as ridge detection [7]. The dark line detecting 
0; vj ∈ Ω1 ∪ ... ∪ Ωk
property of hessian matrix is used to detect these ridges. As wij = (5)
h(vj ); vj ∈ Ω0
single-scale ridge detectors detect only specific width of a
vessel so four different scales are used to detect vessels of Here we have set wij = 0 if vj belongs to foreground,
radii 1 to 7 pixels. An eigenvalues of the hessian matrix for wij = h(vj ) if vj belongs to background. In order to find
image I are calculated as: the shortest path for foreground pixel, weight for pixels
belongs to the foreground is kept to zero. The term h(vj )
depends upon response of ridge detector. Ridge detector has a
 
I Ixy
H = xx (1) maximum response at the center of the vessel. If the response
Ixy Iyy
(a) (b) (c)
Fig. 3. Output of vessel segmentation. (a) Color retinal image from
DRIVE database, (b) initial segmentation results, (c) fully connected Fig. 4. The bifurcation structure and its features.
vascular tree.

Degree of similarity Si,j for bifurcation structure pair is given


is maximized, then the weight of the arc should be minimum. as
Hence h(vj ) is calculated as Si,j = d(pi , qj ) (9)
f (vj ) Where pi and qj are ith and jth bifurcation structure in two
h(vj ) = 1 − (6)
max[f ] images and d is the simple distance between two feature vec-
Where f (vj ) is the output of ridge detector at given vertex tors. Here dummy bifurcation, crossover and terminal points
and h(vj ) has normalized in range [0, 1]. The shortest path are discarded as these points are more error-prone.
follows the pixel having a maximum ridge response. We have 2) Verification of Matched Pairs: Verification of matched
used 8-neighborhood connectivity for defining a graph. A pair is very important to avoid false correspondence. Minimum
single source is selected randomly from largest ridge region. correspondence points required for linear and similarity trans-
This algorithm finds the shortest path from the source to every formation are 2 and 3 pairs respectively. In our method single
other vertex simultaneously. After finding shortest path fully matched bifurcation structure give four point pairs. Therefore,
connected vessel tree is formed by joining these paths. Fig. one matched bifurcation structure is enough for registration.
3(c) shows fully connected vascular tree. Matched pairs are verified as:

C. Image Registration err(ij, mn) = d(M (pi , qj ), M (pm , qn )) (10)


Proposed method of retinal image registration is based on Here M (pi , qj ) and M (pm , qn ) are transformation models
feature matching hence bifurcation structure of blood vessels estimated from matched pairs pi and qj , pm and qn respec-
is used for finding features. The process of forming feature tively. If matched pairs are correct, then the difference between
vector is described in the following section. transformation models of matched pairs should be minimum.
1) Matching of Bifurcation Structure: Bifurcation structure Matched pairs having more difference in transformation model
is composed of one center bifurcation point and is surrounded is discarded in this step.
by three neighbors [9]. Fig. 4 shows the bifurcation structure 3) Validation of Registered Images: Images of different
showing features. Feature vector should be robust to trans- field of view of the same retina contain different overlapping
lation, rotation, scaling, distortion and other transformation. area. A number of matched bifurcation structure varies accord-
Hence we have normalized the feature vector. Bifurcation ing to the overlapping area. Images with more overlapping area
structure consists of four bifurcation points, hence the feature are having more number of structure matches found. There
vector is a combination of a characteristic of these four points. is a need to verify that how many structures are accurately
Feature vector of each bifurcation structure is given as: matched, since for estimating the transformation model true
correspondence is required. This is done by verifying the trans-
V = [lengths, angles] (7) formed image. In registration, one image act as a reference
image and we find transformed second image using n best-
matched structure. Before computing overlapping pixels, both
V = [l1 , l2 , l3 , θ1 , θ2 , θ3 , θ4 , θ5 , θ6 , θ7 , θ8 , θ9 , θ10 , θ11 , θ12 ] images are dilated using the disc structuring element of size
(8) 3×3 to compensate the very small registration error.
Where li and θi are the normalized lengths and angles
respectively. IfTinal = max[(Iref ⊕ S) × (InT ⊕ S)] (11)
li =ith branch length/sum[length1, length2, length3] Iregistered = Iref + IfTinal (12)
θi =ith angle in degree/360o
Feature vector V is arranged in ascending order of lengths Where, Iref is vessel segmented reference image, IT is
and angles. Let P and Q denote feature group of two given vessel segmented transformed second image, S is structuring
images having N1 and N2 bifurcation structures, respectively. element and n is a number of bifurcation structure used for
TABLE I. E XPERIMENTAL R ESULTS
Method Mean Err (pixels) Std. Devi.(pixels)
GDB-ICP 3.505 2.789
Proposed method 1.785 0.974

finding transformation model. Value of n for which maximum


overlapping obtain is used to get transformation of image.
(a) (b)
III. E XPERIMENTAL R ESULTS
Presently there is no public dataset of retinal images avail-
able for retinal image registration. We have done experimenta-
tion using proposed method on a private dataset collected from
Sushrusha Hospital, Nanded. Dataset consists of a different
set of images of individual patients covering different fields
of view images specially collected for this research work. We
have registered images having different pathologies, different
(c) (d)
overlapping areas and low contrast.
Retinal images first segmented using ridge based vessel
segmentation and then fully connected vessel tree are obtained
using a Dijkstra’s algorithm. Proposed retinal image registra-
tion method is applied to fully segmented retinal vascular tree.
We have tried a different transformation model and observed
that similarity transformation model gives the best result. In
Fig. 5, a rectangle marks matched bifurcation point. Here four
best-matched bifurcation structure is shown. Here images are (e) (f)
of size 858×570.
Fig. 5. Result of registration. (a)Reference retinal image, (b)Vascular
Performance of proposed retinal image registration is eval- tree of (a) with matched bifurcation structures, (c)Second retinal
uated using mean and standard deviation of registration error image of same retina, (d)Vascular tree of (c) with matched bifurcation
[1]. Registration error of image pair is given as the mean structures, (e)Mosaic retinal image and (f)Mosaic vescular image.
Euclidian distance between control points of the reference
image and the transformed image. We have manually selected
8 to 20 points of reference image which is also present in the R EFERENCES
second image. These points are bifurcation points as it maps [1] J. Zheng, J. Tian, K. Deng, X. Dai, X. Zhang, and M. Xu, “Salient feature
the whole retina and localized accurately. We have verified region: a new method for retinal image registration,” IEEE Transactions
on Information Technology in Biomedicine, vol. 15, no. 2, pp. 221–232,
the proposed method using six image pairs by comparing its 2011.
result with GDB-ICP algorithm [5] given in Table I. [2] J. Kim and J. A. Fessler, “Intensity-based image registration using robust
correlation coefficients,” IEEE transactions on medical imaging, vol. 23,
IV. C ONCLUSION no. 11, pp. 1430–1444, 2004.
[3] G. K. Matsopoulos, N. A. Mouravliansky, K. K. Delibasis, and K. S.
Proposed retinal image registration method uses vascular Nikita, “Automatic retinal image registration scheme using global opti-
structure to find the correspondence. This method is robust to mization techniques,” IEEE Transactions on Information Technology in
translation, rotation, scaling and intensity of the images. Reg- Biomedicine, vol. 3, no. 1, pp. 47–60, 1999.
[4] C. V. Stewart, C.-L. Tsai, and B. Roysam, “The dual-bootstrap iterative
istered image visually looks very good and also the registration closest point algorithm with application to retinal image registration,”
error is very small. Image preprocessing improves contrast IEEE transactions on medical imaging, vol. 22, no. 11, pp. 1379–1394,
between vessels and background which improves segmenta- 2003.
[5] G. Yang, C. V. Stewart, M. Sofka, and C.-L. Tsai, “Automatic robust
tion. Proposed method removes false matching of structure. A image registration system: Initialization, estimation, and decision,” in
feature based registration method is proposed which uses the Fourth IEEE International Conference on Computer Vision Systems
bifurcation structure for finding the correspondence. Similarity (ICVS’06), pp. 23–23, IEEE, 2006.
[6] N. M. Salem, S. A. Salem, and A. K. Nandi, “Segmentation of retinal
transformation gives the best result. blood vessels based on analysis of the hessian matrix and clustering algo-
rithm,” in Signal Processing Conference, 2007 15th European, pp. 428–
ACKNOWLEDGMENT 432, IEEE, 2007.
[7] J. Staal, M. D. Abràmoff, M. Niemeijer, M. A. Viergever, and B. van
The authors would like to express gratitude toward Dr. Ginneken, “Ridge-based vessel segmentation in color images of the
Girish Deshmukh, Susrusha Hospital, Nanded for support, giv- retina,” IEEE transactions on medical imaging, vol. 23, no. 4, pp. 501–
ing clinical insight and sharing retinal images for our research 509, 2004.
[8] M. Barbehenn, “A note on the complexity of dijkstra’s algorithm for
work. We would also like to thank Technical Education Quality graphs with weighted vertices,” IEEE transactions on computers, vol. 47,
Improvement Program (TEQIP II) for supporting our project. no. 2, p. 263, 1998.
[9] L. Chen, X. Huang, and J. Tian, “Retinal image registration using topo-
logical vascular tree segmentation and bifurcation structures,” Biomedical
Signal Processing and Control, vol. 16, pp. 22–31, 2015.

Das könnte Ihnen auch gefallen