基于深度学习和二维高斯拟合的视网膜血管管径测量方法
|
摘要
糖尿病、高血压等疾病会引起视网膜血管的形状发生变化,眼底图像血管分割是疾病定量分析过程中的关键步骤,对临床疾病的分析和诊断具有指导意义。本文提出一种视网膜血管管径自动测量方法。首先,将通道特征图叠加,同时通过使用深度可分离卷积来增加网络深度,将二者引用于全卷积神经网络中对血管网络进行分割;然后在分割的血管网络基础上,利用形态学细化和最小二乘拟合求取血管的中心线和方向;最后根据血管横截面灰度值分布特性,利用二维高斯拟合对血管中心线和方向进行校正,得到准确的血管方向和中心线位置进而计算血管管径。利用本文方法分别对REVIEW数据库中的3个图像集进行测试,测量的管径均值的标准差接近专家测量的标准差,表明本文血管管径测量方法的准确率高,实验结果验证了本文方法的准确性。
Diseases such as diabetes and hypertension can lead to changes in the shape of retinal blood vessels. Segmentation of fundus images is critical in the quantitative analysis of diseases, which is instructive in the clinical analysis and diagnosis of diseases. Herein a method based on depth learning and two-dimensional Gaussian fitting is proposed to automatically measure the diameters of retinal vessels. Firstly, the connected channel signatures and the network depth which is increased by depthwise separable convolutions are applied into fully convolutional network to segment the vessel network. Then, morphological refinement and least-squares fitting are used to find the centerline and direction of blood vessels in the segmented vessel network.Finally, based on gray value distribution characteristics of blood vessel cross-section, two-dimension Gaussian fitting is used to correct the centerlines and directions of retinal vessels for obtaining the accurate centerlines and directions of retinal vessels and then measuring the diameters of retinal vessels. Three image sets in the REVIEW database is tested. The standard deviation obtained with the proposed method is close to that of the manual measurement, which indicates that the proposed method can achieve a high accuracy in vessel diameter measurement. The validity of the proposed method is proved by the test on the REVIEW database.
Diseases such as diabetes and hypertension can lead to changes in the shape of retinal blood vessels. Segmentation of fundus images is critical in the quantitative analysis of diseases, which is instructive in the clinical analysis and diagnosis of diseases. Herein a method based on depth learning and two-dimensional Gaussian fitting is proposed to automatically measure the diameters of retinal vessels. Firstly, the connected channel signatures and the network depth which is increased by depthwise separable convolutions are applied into fully convolutional network to segment the vessel network. Then, morphological refinement and least-squares fitting are used to find the centerline and direction of blood vessels in the segmented vessel network.Finally, based on gray value distribution characteristics of blood vessel cross-section, two-dimension Gaussian fitting is used to correct the centerlines and directions of retinal vessels for obtaining the accurate centerlines and directions of retinal vessels and then measuring the diameters of retinal vessels. Three image sets in the REVIEW database is tested. The standard deviation obtained with the proposed method is close to that of the manual measurement, which indicates that the proposed method can achieve a high accuracy in vessel diameter measurement. The validity of the proposed method is proved by the test on the REVIEW database.
引文
[1]李满,杨永升,安娜,等.视网膜血管管径测量方法的进展及其影响因素[J].国际眼科杂志, 2014, 14(9):1608-1611.LI M, YANG Y S, AN N, et al. Progress in the measurement of retinalvascular diameter and its influencing factors[J]. International EyeScience, 2014, 14(9):1608-1611.
[2]付新. 2型糖尿病合并高血压病患者尿微量白蛋白临床治疗研究[D].郑州:郑州大学, 2009.FU X. Clinical treatment of microalbuminuria in patients with type 2diabetes and hypertension[D]. Zhengzhou:Zhengzhou University,2009.
[3] LIEW G, SIM D A, KEANE P A, et al. Diabetic macular ischaemia isassociated with narrower retinal arterioles in patients with type 2diabetes[J]. Acta Ophthalmol, 2015, 93(1):45-51.
[4]王淑珍.一种血管管径的精确测量方法[J].中国医学物理学杂志,1999, 16(3):146-147.WANG S Z. An accurate method for measuring vascular diameter[J].Chinese Journal of Medical Physics, 1999, 16(3):146-147.
[5]陆建荣,斯海臣,郁道银.冠状动脉血管直径测量方法的研究[J].中国生物医学工程学报, 2005, 24(4):510-512.LU J R, SI H C, YU D Y. A study of the measurement of coronaryartery diameter[J]. Chinese Journal of Biomedical Engineering,2005, 24(4):510-512.
[6]姚畅,陈后金.一种基于先验知识的视网膜血管宽度测量方法[J].中国科技论文在线, 2009, 4(1):64-68.YAO C, CHEN H J. Measurement of retinal vessel widths based onprior knowledge[J]. Sciencepaper Online, 2009, 4(1):64-68.
[7] ZHOU L, RZESZOTARSKI M S, SINGERMAN L J, et al. Thedetection and quantification of retinopathy using digital angiograms[J]. IEEE Trans Med Imaging, 1994, 13(4):619-626.
[8] LI H, HANDSAKER B, WYSOKER A, et al. The sequence Alignment/Map format and SAMtools[J]. Bioinformatics, 2009, 25(16):2078-2079.
[9] LISKOWSKI P, KRAWIEC K. Segmenting retinal blood vessels withdeep neural networks[J]. IEEE Trans Med Imaging, 2016, 35(11):2369-2380.
[10]FU H, XU Y, WONG D W, et al. Retinal vessel segmentation via deeplearning network and fully-connected conditional random fields[C]//IEEE International Symposium on Biomedical Imaging. IEEE, 2016:698-701.
[11]SHELHAMER E, LONG J, DARRELL T. Fully convolutionalnetworks for semantic segmentation[J]. IEEE Trans Pattern AnalMach Intell, 2017, 39(4):640-651.
[12]RONNEBERGER O, FISCHER P, BROX T. U-Net:convolutionalnetworks for biomedical image segmentation[C]//Computer Visionand Pattern Recognition, 2015, 9351:234-241.
[13]DECO G, SCARANO L, SOTO-FARACO S. Weber's law in decisionmaking:integrating behavioral data in humans with aneurophysiological model[J]. J Neurosci, 2007, 27(42):11192-11200.
[14]FAWCETT T. An introduction to ROC analysis[J]. Pattern RecognitiLett, 2006, 27(8):861-874.
[15]ZHANG J, DASHTBOZORG B, BEKKERS E, et al. Robust retinalvessel segmentation via locally adaptive derivative frames inorientation scores[J]. IEEE Trans Med Imaging, 2016, 35(12):2631-2644.
[16]LI Q, FENG B, XIE L P, et al. A cross-modality learning approach forvessel segmentation in retinal images[J]. IEEE Trans Med Imaging,2016, 35(1):109-118.
[17]WANG Y, JI G, LIN P, et al. Retinal vessel segmentation usingmultiwavelet kernels and multiscale hierarchical decomposition[J].Pattern Recognit, 2013, 46(8):2117-2133.
[18]ZHAO Y Q, WANG X H, WANG X F, et al. Retinal vesselssegmentation based on level set and region growing[J]. PatternRecognit, 2014, 47(7):2437-2446.
[19]AZZOPARDI G, STRISCIUGLIO N, VENTO M, et al. TrainableCOSFIRE filters for vessel delineation with application to retinalimages[J]. Med Image Anal, 2015, 19(1):46-57.
[20]ZHOU L, RZESZOTARSKI M S, SINGERMAN L J, et al. Thedetection and quantification of retinopathy using digital angiograms[J]. IEEE Trans Med Imaging, 1994, 13(4):619-626.
[21]LOWELL J, HUNTER A, STEEL D, et al. Quantitative measurementof retinal vessel widths from fundus images based on 2D modelling[J]. IEEE Trans Med Imaging, 2004, 23(10):1196-1204.
[22]LUPA?CU C A, TEGOLO D, TRUCCO E. Accurate estimation ofretinal vessel width using bagged decision trees and an extendedmultiresolution Hermite model[J]. Med Image Anal, 2013, 17(8):1164-1180.
[23]BRINCHMANN-HANSEN O, HEIER H. Theoretical relationsbetween light streak characteristics and optical properties of retinalvessels[J]. Acta Ophthalmol, 1986, 64(S179):33-37.
[24]AL-DIRI B, HUNTER A, STEEL D. An active contour model forsegmenting and measuring retinal vessels[J]. IEEE Trans MedImaging, 2009, 28(9):1488-1497.
[2]付新. 2型糖尿病合并高血压病患者尿微量白蛋白临床治疗研究[D].郑州:郑州大学, 2009.FU X. Clinical treatment of microalbuminuria in patients with type 2diabetes and hypertension[D]. Zhengzhou:Zhengzhou University,2009.
[3] LIEW G, SIM D A, KEANE P A, et al. Diabetic macular ischaemia isassociated with narrower retinal arterioles in patients with type 2diabetes[J]. Acta Ophthalmol, 2015, 93(1):45-51.
[4]王淑珍.一种血管管径的精确测量方法[J].中国医学物理学杂志,1999, 16(3):146-147.WANG S Z. An accurate method for measuring vascular diameter[J].Chinese Journal of Medical Physics, 1999, 16(3):146-147.
[5]陆建荣,斯海臣,郁道银.冠状动脉血管直径测量方法的研究[J].中国生物医学工程学报, 2005, 24(4):510-512.LU J R, SI H C, YU D Y. A study of the measurement of coronaryartery diameter[J]. Chinese Journal of Biomedical Engineering,2005, 24(4):510-512.
[6]姚畅,陈后金.一种基于先验知识的视网膜血管宽度测量方法[J].中国科技论文在线, 2009, 4(1):64-68.YAO C, CHEN H J. Measurement of retinal vessel widths based onprior knowledge[J]. Sciencepaper Online, 2009, 4(1):64-68.
[7] ZHOU L, RZESZOTARSKI M S, SINGERMAN L J, et al. Thedetection and quantification of retinopathy using digital angiograms[J]. IEEE Trans Med Imaging, 1994, 13(4):619-626.
[8] LI H, HANDSAKER B, WYSOKER A, et al. The sequence Alignment/Map format and SAMtools[J]. Bioinformatics, 2009, 25(16):2078-2079.
[9] LISKOWSKI P, KRAWIEC K. Segmenting retinal blood vessels withdeep neural networks[J]. IEEE Trans Med Imaging, 2016, 35(11):2369-2380.
[10]FU H, XU Y, WONG D W, et al. Retinal vessel segmentation via deeplearning network and fully-connected conditional random fields[C]//IEEE International Symposium on Biomedical Imaging. IEEE, 2016:698-701.
[11]SHELHAMER E, LONG J, DARRELL T. Fully convolutionalnetworks for semantic segmentation[J]. IEEE Trans Pattern AnalMach Intell, 2017, 39(4):640-651.
[12]RONNEBERGER O, FISCHER P, BROX T. U-Net:convolutionalnetworks for biomedical image segmentation[C]//Computer Visionand Pattern Recognition, 2015, 9351:234-241.
[13]DECO G, SCARANO L, SOTO-FARACO S. Weber's law in decisionmaking:integrating behavioral data in humans with aneurophysiological model[J]. J Neurosci, 2007, 27(42):11192-11200.
[14]FAWCETT T. An introduction to ROC analysis[J]. Pattern RecognitiLett, 2006, 27(8):861-874.
[15]ZHANG J, DASHTBOZORG B, BEKKERS E, et al. Robust retinalvessel segmentation via locally adaptive derivative frames inorientation scores[J]. IEEE Trans Med Imaging, 2016, 35(12):2631-2644.
[16]LI Q, FENG B, XIE L P, et al. A cross-modality learning approach forvessel segmentation in retinal images[J]. IEEE Trans Med Imaging,2016, 35(1):109-118.
[17]WANG Y, JI G, LIN P, et al. Retinal vessel segmentation usingmultiwavelet kernels and multiscale hierarchical decomposition[J].Pattern Recognit, 2013, 46(8):2117-2133.
[18]ZHAO Y Q, WANG X H, WANG X F, et al. Retinal vesselssegmentation based on level set and region growing[J]. PatternRecognit, 2014, 47(7):2437-2446.
[19]AZZOPARDI G, STRISCIUGLIO N, VENTO M, et al. TrainableCOSFIRE filters for vessel delineation with application to retinalimages[J]. Med Image Anal, 2015, 19(1):46-57.
[20]ZHOU L, RZESZOTARSKI M S, SINGERMAN L J, et al. Thedetection and quantification of retinopathy using digital angiograms[J]. IEEE Trans Med Imaging, 1994, 13(4):619-626.
[21]LOWELL J, HUNTER A, STEEL D, et al. Quantitative measurementof retinal vessel widths from fundus images based on 2D modelling[J]. IEEE Trans Med Imaging, 2004, 23(10):1196-1204.
[22]LUPA?CU C A, TEGOLO D, TRUCCO E. Accurate estimation ofretinal vessel width using bagged decision trees and an extendedmultiresolution Hermite model[J]. Med Image Anal, 2013, 17(8):1164-1180.
[23]BRINCHMANN-HANSEN O, HEIER H. Theoretical relationsbetween light streak characteristics and optical properties of retinalvessels[J]. Acta Ophthalmol, 1986, 64(S179):33-37.
[24]AL-DIRI B, HUNTER A, STEEL D. An active contour model forsegmenting and measuring retinal vessels[J]. IEEE Trans MedImaging, 2009, 28(9):1488-1497.