This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2014.2365514, IEEE Journal of Biomedical and Health Informatics

> JBHI-00037-2014
JBHI-00037-2014
JBHI-00037-2014
JBHI-00037-2014
JBHI-00037-2014
JBHI-00037-2014 < orientation of main arcade of the vasculature. By observation, we believe p ∈ [9,81] is an appropriate range for our parabola fitting algorithm. The range of yOD is the integer between 1 and image height. However, because the OD center is the convergence point of blood vessels, it is impossible present in vessel periphery region, so the lower and up bound of vessel vertical coordinate at position xOD can be adopted to confine the possible search range of yOD (Fig. 3(a)). To accelerate search process, the step great than 1 is appreciated to identify yOD . In our experiments, this step parameter is set to 2, which accelerates the efficiency of the algorithm, as well as doesn’t significantly reduce the localization precision of OD center. To reduce the search space, we expect to implement parabola fitting with single pixel width vessel structure, therefore, after Gabor filtering, a non-maxima suppression thinning technology [8] is used to get central pixels of blood vessels. Due to the main arcade of the vessel structure provide enough global vessel directional information, thus we only need the main arcade to achieve OD vertical localization. To extract the main arcade of vasculature, the vessel bifurcations approximate vertical direction ( > 70 0 ) or less than 20 pixels are removed (Fig. 4(a)). To improve the efficiency of follow-up parabola fitting, whose computation cost is greatly depended on the number of detected vessel pixels, a blocking technology is adopted to

6 simplify the main arcade of vessel structure, and it reduces the time for OD vertical localization by GHT parabola fitting. To achieve blocking vessel map, the image space is split into 10×10 not overlapped blocks, and if there are vessel pixels fall into a block, then it will be filled with “1”, else it will be filled with “0”, therefore a blocking binary vessel map can be acquired(Fig. 4(c)). Here each white block will be used to fitting the parabolas by GHT transformation. In Fig. 4(a), there are 3446 vessel pixels need to be fitted, while in Fig. 4(c), there are only 427 white blocks should be explored. Apparently, the computation time will be greatly reduced by our blocking technology. More than one parabola will be found by above GHT algorithm. The one fitting most of vessel pixels and according best with the global vessel direction need to be found from all possible parabolas. At specific k candidate OD horizontal positions, the parabola that fitting most of vessel blocks are chosen in advance, then the orientation of vessel pixels locating on k specific parabola curves are extracted to evaluate the degree of the parabola according with global vessel direction. The direction of any point ( x, y ) belonging to parabola curve can be given by following equation ⎛ ⎞ 2p θ mod ( x, y ) = arctan ⎜ (9) ⎟ ⎝ −( y − yOD ) ⎠ And the direction of vessel pixels can be determined by Gabor filter with equation,

 

(a) Thinning main arcade of vessel map

(b) Horizontal projection curve of f D ( x)

 

(c) Blocking vessel map

(d) 5 parabolas fitting most of vessel pixels at 5 candidate xOD

Fig. 4. The illustration of OD vertical localization.

2168-2194 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2014.2365514, IEEE Journal of Biomedical and Health Informatics

> JBHI-00037-2014
JBHI-00037-2014 < vessel structure. To evaluate the effect of vessel structure for OD localization, TABLE V and TABLE VI show the number of failed images during OD detection by our method for DRIVE and STARE image dataset. Coarse to finer vasculature are presented with 5% ~ 20% FOV vessel result. And k (k=1, 3, 5) minimum extreme points are selected as possible horizontal coordinate of OD. Obviously, as more vessel details are presented, the failed images reduce, however if there are much false positive vessel pixels, the failed images may be increase, e.g. for DRIVE dataset, there are 2 failed images in 20% FOV vessel result when k=1, while all images can be correctly detected with 15% FOV vessel results and k=1. Also for STARE dataset, there are 5 or 6 failed images with k=1 and 15% or 20% FOV vessel results respectively, while there are only 4 failed images with 10% and 12% FOV vessel results. Moreover, we can observe that in every vessel results, the failed images will decrease with the increase of k. It means that when appropriate vasculature is extracted, the robust of the method can be achieved when a certain amount of xOD candidates are selected. Although the method show excellent OD detection accuracy, however it should be noticed that complete and accurate vasculature is still favor to OD detection, the best OD detection accuracy is achieved when manual labeled vessel are used to OD localization for our method in both DRIVE and STARE image datasets. But as the increase of k, e.g. k=5, our method also can achieve best OD accuracy under 8%~20% vessel results. Due to too much vessel pixels information are lost, 5% FOV vessel results present worst OD accuracy. The OD detection accuracy of 4 public image datasets with our method is listed in TABLE VII. The best accuracy (99.7%) is achieved when k=5 and 10% or 12% vessel results are provided. Moreover the accuracy has no significant variation when 8% ~20% vessel results are used to localize the OD region in specified k. We observe that the variation of the accuracy is no more than 2%. It shows that our method is robust to vessel structure and exhibits no great dependence on finer vessel detection results. By experiment, we ultimately choose k=5 and 10% vessel results as the final parameters to our algorithm. There are total of 340 images in 4 datasets. And only 1 image in STARE is failed. We know that there are many retinal images suffer from different lesions and imaging artifacts in these test images. Thus our proposed method presents excellent OD detection accuracy (99.7%). And it is robust to various lesions and different qualities retinal images. STARE dataset includes many pathological images and widely used as a benchmarking in much of state-of-the-art OD detection methods. TABLE VIII shows the OD detection accuracy of our method and other state-of-the-art methods. Lu’s[6], Youssif’s[3] and our method present best accuracy(98.8%), and only 1 image is failed. Many OD detection literatures mentioned the speed of their algorithms. TABLE VIII lists the reported times of other algorithms in the original papers [1-6]. The most efficient method presently reported is image feature projection approach [5]. Due to

8 reducing the 2D problem into two 1D projection problems, it achieved OD detection in a STARE image with only 0.46s. The best accuracy of STARE dataset has been reported is Youssif [3] and Lu [6], they can achieve 98.8% accuracy in 3.5min and 5s, respectively. Obviously 3.5min is too much to real application. We also notice that the speed of Lu’s [6] is achieved in a small size of original retinal image, i.e., 0.3 of its original size. We observe that the vasculature is the most stable feature of retinal image, and it also presents scale invariant characteristic. Thus it is reasonable believed that we can achieve OD detection with higher speed in small size retinal image. To evaluate the performance of our method in different scale retinal images, we resize the STARE dataset with 4 scales, i.e., 1, 0.7, 0.5, 0.3 of its original size. TABLE VIII OD DETECTION RESULTS FOR THE PROPOSED AND LITERATURE REPORTED METHODS IN STARE DATASET #of failed Methods Accuracy Speed image 3 4.5min Lu's [4] 96.3% 1 5s Lu's [6] 98.8% 9 15s Hoover [2] 89.0% 2 2 min Foracchia [1] 97.5% 1 3.5 min Youssif [3] 98.8% 6 0.46 s Mahfouz [5] 92.6% 1 3.4-11.5s The proposed 98.8%

TABLE IX shows accuracy of STARE dataset in different scales. And it presents excellent robust against scale variation. i.e., the accuracy has no decrease with the change of image size, while the efficiency is greatly improved (from 11.5s to 3.4s). In experiment, the algorithm is implemented using MATLAB 2010 on Windows XP SP3 with an Intel Core 2 Duo P8400 2.4GHz CPU with RAM DDR2 2GB. TABLE IX OD DETECTION RESULTS OF PROPOSED METHOD FOR STARE DATASET WHEN RESIZED IMAGES WITH DIFFERENT SCALES ARE TESTED Scale Accuracy Speed 1 98.8% 11.5s 0.7 98.8% 6.5 s 0.5 98.8% 4.2 s 0.3 98.8% 3.4 s

Fig. 5 shows OD detection result of 12 pathological images by our method. In each image, the green parabola curve is the final parabola we find to identify OD center, and the vertex of the parabola, i.e. OD center, is marked with red “+”. All these images are chosen from STARE dataset and suffer from much pathology or imaging artifacts. Among them, only Fig. 5(j) is failed, i.e., the OD center we find is fall outside of the OD boundary, although it is very close to the OD boundary. Considering the vessel distribution and direction characteristics are applied to detect the OD location in our algorithm, thus we believe that the main reason for the failure was the image im0041 can’t supply a sizable vessel structure. IV. DISCUSSION The proposed method presents high OD detection accuracy

2168-2194 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JBHI.2014.2365514, IEEE Journal of Biomedical and Health Informatics

> JBHI-00037-2014
JBHI-00037-2014

Novel Accurate and Fast Optic Disc Detection in Retinal Images With Vessel Distribution and Directional Characteristics.

A novel accurate and fast optic disc (OD) detection method is proposed by using vessel distribution and directional characteristics. A feature combini...
523KB Sizes 0 Downloads 4 Views