Вы находитесь на странице: 1из 4

Retinal Vessel Segmentation Based on Frangi Filter

and Morphological Reconstruction

Hanung Adi Nugroho*, Rezty Amalia Aras, Tri Lestari, Igi Ardiyanto
Department of Electrical Engineering and Information Technology
Faculty of Engineering, Universitas Gadjah Mada
Jalan Grafika 2, Kampus UGM, Yogyakarta 55281, Indonesia
*adinugroho@ugm.ac.id

Abstract—The analysis of structural changes in retinal


vessels is the most important part for diagnosing and
detecting retinal related diseases such as diabetic
retinopathy, hypertension, age-related macular
degeneration (AMD) and arteriosclerotic. This paper Retinal vessels

presents a method for segmenting retinal vessels in retinal


Macula
fundus image based on Frangi filter and morphological
reconstruction. The proposed method is evaluated using
colour fundus images from DRIVE and STARE datasets.
In DRIVE dataset, the performance of proposed method Optic disc Foveal avascular zone
(FAZ)
achieves an average of sensitivity, specificity and accuracy
at 72.13%, 96.65 % and 94.50%, respectively. Meanwhile,
in STARE dataset, it achieves an average sensitivity of
75.50%, specificity of 90.38% and accuracy of 88.76%.
These results indicate that the proposed method
successfully segments retinal vessels in fundus images. Fig. 1. The main structure of retinal image [7]
Keywords—retinal related diseases; retinal vessels; Frangi filter;
morphological reconstruction; fundus image In this research work, a method for segmentation of retinal
blood vessels from retinal fundus images is proposed by
I. INTRODUCTION combining Frangi filter with morphological reconstruction.
In ophthalmology, detection and segmentation of blood The proposed method is applied in DRIVE and STARE
vessels are essentially required as part of analysis of colour datasets. The paper is written in several sections. Having
retinal images for monitoring and diagnosis the severity of introduced the problems in first section, the second section
retinal related diseases, such as diabetic retinopathy, describes the dataset and the proposed scheme. In the third
hypertension, AMD and arteriosclerotic [1, 2]. Early detection section, the results are presented and analysed. Subsequently,
in diagnosing retinal related diseases is important to give the conclusion is presented in the final section.
appropriate treatment and to prevent vision loss [3]. Various
structures and components are present in the human retina such II. DATASET AND METHOD
as foveal avascular zone (FAZ), macula, optic disc and retinal This study uses a total of sixty digital colour retinal images
blood vessels as shown in Fig. 1. Early symptoms of disease consisting of forty images from DRIVE dataset [7] and twenty
are expressed by features of retinal blood vessels such as images from STARE dataset [8]. Ground truth images are also
diameter, length, angle and tortuosity [4]. Segmentation of available in these datasets. The proposed method consists of
retinal blood vessels can facilitate the ophthalmologists in two main schemes. The two schemes conducted in this study
diagnosing eye diseases as well as clinical studies and screening are illustrated in Fig. 2.
for patients.
A. Frangi Filter for Vessels Detection
Several methods of retinal blood vessels segmentation have
been developed in previous studies. Quinn et al. [5] proposed The Frangi filter can be used to detect continuous edges,
retinal blood vessels segmentation using curvelet transform and such as vessels. It uses eigenvalues of Hessian matrix to find
morphological reconstruction, while Sasidharan et al. [6] tubular structures. In Frangi method, 2D pixels are expressed
presented the extraction and segmentation of coronary arteries in 3D by adding intensity values of Z-coordinate. The equation
from X-ray angiographic images using noise adaptive fuzzy is formulated in (1).
switching median (NAFSM) filter combined with Frangi’s
vessel detection and region growing. 𝐶 = (𝑥, 𝑦, 𝐼) (1)
with I is pixel intensity (x, y). Hessian matrix is the second Operation of morphological reconstruction is a continuation
order partial derivative of a function as formulated in (2). of pre-processing stage, such as extraction of green channel,
complement of green channel and CLAHE. Each of this
𝐼𝑥𝑥 (𝑥, 𝑦) 𝐼𝑥𝑦 (𝑥, 𝑦) process is described as the following.
𝐻=( ) (2)
𝐼𝑦𝑥 (𝑥, 𝑦) 𝐼𝑦𝑦 (𝑥, 𝑦) C. Extraction of Green Channel
Based on the histogram of red, green and blue channels as
Two Eigen values can be calculated for each pixel from this shown in Fig. 3, the green channel shows the best contrast
Hessian matrix. To extract small vessels or known as retinal between retinal vessels and the background. On the other hand,
capillaries, this operation is applied on different scales. The red and blue channels have more noises. Therefore, the green
maximum of V is calculated across multiple scales for getting channel is selected to minimise the presence of noise as well as
final result as mathematically formulated in (3). to obtain more enhanced of retinal blood vessels.
𝐹(𝑥, 𝑦) = 𝑚𝑎𝑥{𝑉(𝑥, 𝑦, 𝑒1 (𝜎))} (3)

with 𝜎 is the scale.

Start

RGB fundus
image
(a) (b)
Scheme-1 Scheme-2
Green channel
Frangi filtering
extraction
Complement
operation

CLAHE

(c) (d)
Median filtering
Fig. 3. The images result of channel extraction (a) original image (b) red
Morphological Morphological channel (c) green channel and (d) blue channel.
reconstruction reconstruction
Morphological Morphological D. Complement Operation
OR Operation
image image
Complement of green channel is used to raise the
Detected illumination of retinal vessels from the digital colour retinal
retinal vessels fundus image. The function of the complement is mentioned in
(6) and (7).
End
𝐴𝑐 = 𝑈\𝐴 (6)
𝑐
Fig. 2. Flowchart of the proposed method 𝐴 = {𝜔 ∈ 𝑈|𝜔 ∉ 𝐴} (7)

B. Morphological Reconstruction with 𝐴𝑐 is the symbol of complement and 𝜔 is the element of


𝑈 or not element of 𝐴. The complement makes the appearance
Reconstruction is a morphological transformation involving of retinal blood vessels brighter than other features in the colour
two images (i.e. the mask and the marker) and a structuring retinal images.
element. If g is the mask and f is the marker, reconstruction of
g from f is declared as Rg(f) which is defined by the following E. Contrast Limited Adaptive Histogram Equalisation
iterative procedure. (CLAHE)
• Initialise ℎ1 into image marker f CLAHE technique is extremely used to enhance the contrast
• Create structuring element: B = ones (3) of medical images. This method operates on the small region
• Repeat of image called ‘tile’. Enhancement calculation is modified
based on contrast-limit level according to user-specified
ℎ𝑘+1 = (ℎ𝑘 ⊕ 𝐵) ∩ 𝑔 (4) maximum by placing limit l, 0 ≤ l ≤ 1.
to ℎ𝑘+1 = ℎ𝑘 F. Median Filter
Median filtering has a strong ability to eliminate noises such
The image marker 𝑓 has to be part of 𝑔; so,
as salt and pepper noise which appears as white and black dots
in the images. This filter replaces a pixel value in the centre
𝑓⊆𝑔 (5)
coordinate of sliding windows with the median value of pixels 𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑥100% (10)
in the window. 𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁
III. RESULTS AND DISCUSSION
A. Scheme-1
In [6], vessels detection based on Frangi filter was
conducted by extracting and segmenting coronary arteries from
X-ray angiographic images. Similarly, in this work, Frangi
filter is applied for segmenting retinal vessels on digital colour
fundus images. As shown in Fig. 4 (b), the image resulted by
Frangi filter still detects the retinal boundary. Therefore, it is (a) (b)
overcome by applying morphological reconstruction as
presented in Fig. 4 (c).

(c) (d)

(a) (b) (c)


Fig. 4. The images result of (a) original RGB image (b) Frangi filter and
(c) morphological reconstruction.

B. Scheme-2 (e) (f)


Channel extraction aims to extract the red, green and blue Fig. 5. The images result of (a) original RGB image (b) green channel
extraction (c) complement operation (d) CLAHE (e) median filtering (f)
channels. The extracted green channel is then selected since it
morphological reconstruction.
has enough information for segmenting retinal blood vessels.
Then, complement operation is applied followed by CLAHE
method to make retinal vessels more clearly. However, this
process yields some noises in the enhanced image. Thus,
median filter is conducted to eliminate these noises. Finally,
filtered image is processed by morphological reconstruction to
obtain the candidates of segmented retinal vessels. The resulted
images of scheme-2 are depicted from Fig. 5 (a) to Fig. 5 (f).
Scheme-1 is able to segment the retinal vessels more clearly;
however, the small vessels tend to be lost. While the scheme-2
successfully segments the small vessels. Therefore, to obtain
better results, retinal vessels segmented from scheme-1 (Fig. 4
(c)) are combined with that of scheme-2 (Fig. 5 (f)) by
employing OR operation. An example of final result of retinal
vessels segmentation is shown in Fig. 6.
To evaluate the performance of the proposed method, two Fig. 6. The final result of segmented retinal vessels
datasets consisting of DRIVE and STARE are used. The retinal
vessels segmented by the proposed method are then compared True positive (TP) is the number of vessels pixels correctly
to the ground truth image. Some parameters such as true segmented as vessels while true negative (TN) is the number of
positive (TP), true negative (TN), false positive (FP) and false non-vessels correctly segmented as non-vessels. False positive
negative (FN) are considered to calculate the rate of sensitivity, (FP) is the number of non-vessels pixels segmented as vessels
specificity and accuracy by using (8) until (10). while false negative (FN) is the number of vessels pixels
segmented as non-vessels. The evaluation results are presented
𝑇𝑃 in Table 1.
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = 𝑥100% (8)
Based on Table 1, implementation of the proposed method
𝑇𝑃 + 𝐹𝑁
𝑇𝑁 on DRIVE dataset obtains the average of sensitivity, specificity
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = 𝑥100% (9) and accuracy at 72.13%, 96.65% and 94.50%, respectively.
𝑇𝑁 + 𝐹𝑃 While the implementation of the proposed method on STARE
dataset achieves the average sensitivity of 75.50%, specificity REFERENCES
of 90.38% and accuracy of 88.76%. [1] A. Mehrotra, S. Tripathi, K. K. Singh, and P. Khandelwal, "Blood Vessel
Extraction for retinal images using morphological operator and KCN
TABLE I. THE EVALUATION RESULTS OF THE PROPOSED METHOD clustering," in Advance Computing Conference (IACC), 2014 IEEE
International, 2014, pp. 1142-1146.
Dataset Sensitivity (%) Specificity (%) Accuracy (%) [2] J. J. Kanski and B. Bowling, Clinical ophthalmology: a systematic
DRIVE 72.13 96.65 94.50 approach: Elsevier Health Sciences, 2011.
[3] N. P. Singh, R. Kumar, and R. Srivastava, "Local entropy thresholding
STARE 75.50 90.38 88.76 based fast retinal vessels segmentation by modifying matched filter," in
2015 International Conference on Computing, Communication &
Automation (ICCCA), 2015, pp. 1166-1170.
IV. CONCLUSION [4] M. E. Martinez-Perez, A. Highes, A. V. Stanton, S. A. Thorn, N.
A method for segmenting retinal vessels based on Frangi Chapman, A. A. Bharath, et al., "Retinal vascular tree morphology: a
semi-automatic quantification," IEEE Transactions on Biomedical
filter and morphological reconstruction has been proposed. The Engineering, vol. 49, pp. 912-917, 2002.
proposed method is validated on DRIVE and STARE datasets. [5] E. A. E. Quinn and K. G. Krishnan, "Retinal blood vessel segmentation
The evaluation results on the DRIVE dataset achieve the using curvelet transform and morphological reconstruction," in 2013
average sensitivity of 72.13 %, specificity of 96.65 % and International Conference on Emerging Trends in Computing,
Communication and Nanotechnology (ICE-CCN), 2013, pp. 570-575.
accuracy of 94.50%. Whilst that of STARE dataset achieves [6] G. Sasidharan and A. George, "Frangi's vessel detection approach for
the average of sensitivity, specificity and accuracy at 75.50%, coronary angiogram segmentation," Int. J Eng. Trends Technol., vol. 13,
90.38 % and 88.76%, respectively. These successful results in pp. 213-217, 2014.
segmenting retinal vessels indicate that the proposed scheme [7] J. Staal, M. D. Abràmoff, M. Niemeijer, M. A. Viergever, and B. Van
Ginneken, "Ridge-based vessel segmentation in color images of the
has a potential to be implemented in the development of retina," IEEE transactions on medical imaging, vol. 23, pp. 501-509,
computerised retinal related diseases detection system. 2004.
[8] Structured Analysis of the Retina [Online]. Available:
ACKNOWLEDGMENT http://cecas.clemson.edu/~ahoover/stare/
The authors would like to acknowledge the Directorate
General of Higher Education, Ministry of Research, Technology
and Higher Education, Republic of Indonesia for financially
supported by the Research Grant “Penelitian Unggulan
Perguruan Tinggi” Universitas Gadjah Mada, No.
2365/UN1.P.III/DIT-LIT/LT/2017. The authors would like to
thank the Intelligence Systems research group members for
inspiring discussion and sharing meaningful knowledge.

Вам также может понравиться