Saroj Sushil Kumar, Kumar Rakesh, Singh Nagendra Pratap
Department of Computer Science and Engineering, MMM University of Technology, Gorakhpur, India.
Department of Computer Science and Engineering, National Institute of Technology, Hamirpur, India.
Comput Methods Programs Biomed. 2020 Oct;194:105490. doi: 10.1016/j.cmpb.2020.105490. Epub 2020 Jun 5.
Retinal pathology diseases such as glaucoma, obesity, diabetes, hypertension etc. have deadliest impact on life of human being today. Retinal blood vessels consist of various significant information which are helpful in detection and treatment of these diseases. Therefore, it is essential to segment these retinal vessels. Various matched filter approaches for segmentation of retinal blood vessels are reported in the literature but their kernel templates are not appropriate to vessel profile resulting poor performance. To overcome this, a novel matched filter approach based on Fréchet probability distribution function has been proposed.
Image processing operations which we have used in the proposed approach are basically divided into three major stages viz; pre processing, Fréchet matched filter and post processing. In pre processing, principle component analysis (PCA) method is used to convert color image into grayscale image thereafter contrast limited adaptive histogram equalization (CLAHE) is applied on obtained grayscale to get enhanced grayscale image. In Fréchet matched filter, exhaustive experimental tests are conducted to choose optimal values for both Fréchet function parameters and matched filter parameters to design new matched filter. In post processing, entropy based optimal thresholding technique is applied on obtained MFR image to get binary image followed by length filtering and masking methods are applied to generate to a clear and whole vascular tree.
For evaluation of the proposed approach, quantitative performance metrics such as average specificity, average sensitivity and average accuracy and root mean square deviation (RMSD) are computed in the literature. We found the average specificity 97.24%, average sensitivity 72.78%, average accuracy 95.09% for STARE dataset while average specificity 97.61%, average sensitivity 73.07%, average accuracy 95.44% for DRIVE dataset. Average RMSD values are found 0.07 and 0.04 for STARE and DRIVE databases respectively.
From experimental results, it can be observed that our proposed approach outperforms over latest and prominent works reported in the literature. The cause of improved performance is due to better matching between vessel profile and Fréchet template.
青光眼、肥胖症、糖尿病、高血压等视网膜病理疾病对当今人类生活有着致命影响。视网膜血管包含各种重要信息，有助于这些疾病的检测与治疗。因此，对这些视网膜血管进行分割至关重要。文献中报道了各种用于视网膜血管分割的匹配滤波方法，但其核模板与血管轮廓不匹配，导致性能不佳。为克服这一问题，提出了一种基于弗雷歇概率分布函数的新型匹配滤波方法。
我们在所提出的方法中使用的图像处理操作主要分为三个主要阶段，即预处理、弗雷歇匹配滤波和后处理。在预处理中，使用主成分分析（PCA）方法将彩色图像转换为灰度图像，然后对得到的灰度图像应用对比度受限自适应直方图均衡化（CLAHE）以获得增强的灰度图像。在弗雷歇匹配滤波中，进行详尽的实验测试以选择弗雷歇函数参数和匹配滤波参数的最佳值，从而设计新的匹配滤波器。在后处理中，对得到的MFR图像应用基于熵的最优阈值技术以获得二值图像，随后应用长度滤波和掩膜方法生成清晰完整的血管树。
为评估所提出的方法，文献中计算了平均特异性、平均敏感性、平均准确性和均方根偏差（RMSD）等定量性能指标。我们发现，对于STARE数据集，平均特异性为97.24%，平均敏感性为72.78%，平均准确性为95.09%；而对于DRIVE数据集，平均特异性为97.61%，平均敏感性为73.07%，平均准确性为95.44%。对于STARE和DRIVE数据库，平均RMSD值分别为0.07和0.04。
从实验结果可以看出，我们提出的方法优于文献中报道的最新且突出的研究成果。性能提升的原因在于血管轮廓与弗雷歇模板之间的更好匹配。
