·1250 工程科学学报，第43卷，第9期 85/a segmentation and registration methods based on supervised 08 learning.Chin J Eng,2020,42(10):1362 E83 Singel task network (丛明，吴童，刘冬，等.基于监督学习的前列腺MR/TRUS图像分 82 割和配准方法.工程科学学报，2020,42(10)：1362) 1234567891011121314151617181920 [2] Training iteration Ma B Y,Jiang S F,Yin D,et al.Image segmentation metric and its application in the analysis of microscopic image.Chin J Eng, 82.5(b) 82.0 2021,43(1):137 (马博渊，姜淑芳，尹豆，等.图像分割评估方法在显微图像分析 80.0 二 中的应用.工程科学学报，2021,43(1)：137) 90 [3] 234567891011121314151617181920 Tso M O M,Jampol L M.Pathophysiology of hypertensive Training iteration retinopathy.Ophthalmology,1982.89(10):1132 图5消融实验中每轮训练后在不同验证集上的F1值.(a)DRIVE: [4] Yu S,Xiao D,Kanagasingam Y.Machine learning based (b)CHASEDBI automatic neovascularization detection on optic disc region./EEE Fig.5 Fl on the validation set after each training iteration in ablation J Biomed Heal Inform,2018,22(3):886 experiments:(a)DRIVE;(b)CHASEDBI [5] Becker C,Rigamonti R,Lepetit V,et al.Supervised feature leamning for curvilinear structure segmentation /International 表3第二组消融实验结果 Conference on Medical Image Computing and Computer-Assisted Table 3 Results of the second ablation experiments Intervention.Berlin,2013:526 DRIVE CHASEDB Network [6]Tolias Y A.Panas S M.A fuzzy vessel tracking algorithm for F1/%Se/%Acc/%AUC/%F1/%Se/%Acc/%AUC/% retinal images based on fuzzy clustering.IEEE Trans Med ResNet3483.183.797.198.882.084.597.199.1 1 naging,1998,17(2):263 ResNet1882.983.597.098.781.783.897.6 99.0 [7] Soares J V B,Leandro JJ G,Cesar R M,et al.Retinal vessel VGG1682.883.2 97.0 98.7 81.683.797.6 99.0 segmentation using the 2-D Gabor wavelet and supervised classification.IEEE Trans Med Imaging,006,25(9):1214 035 [8] 0.30 Sebbe R,Gosselin B,Coche E,et al.Segmentation of opacified (a) 025 3020 thorax vessels using model-driven active contour //2005 IEEE 90.15 Engineering in Medicine and Biology 27th Annual Conference 0.10 0.05 Shanghai,2006:2535 0 1234567891011121314151617181920 [9] Pal S,Chatterjee S,Dey D,et al.Morphological operations with Epoch iterative rotation of structuring elements for segmentation of 0.35 0.30 6 retinal vessel structures.Multidimens Syst Signal Process,2019, 025 30(1):373 0.20 0.15 [10]Chang C C,Lin CC,Pai P Y,et al.A novel retinal blood vessel 0.10 0.05 segmentation method based on line operator and edge detector// 0 1234567891011121314151617181920 2009 Fifth International Conference on Intelligent Information Epoch Hiding and Multimedia Signal Processing.Kyoto,2009:299 图6框架在DRIVE(a)和CHASEDBI(b)数据集上的训练损失 [11]Zhang Y S,Chung A C S.Deep supervision with additional labels Fig.6 Training loss of the framework on the DRIVE (a)and for retinal vessel segmentation task /International Conference on CHASEDB1 (b)datasets Medical Image Computing and Computer-Assisted Intervention. 4结论 Granada,2018:83 [12]Ronneberger O,Fischer P,Brox T.U-net:Convolutional networks 本文提出了一种骨架图引导的多任务级联视 for biomedical image segmentation /International Conference on 网膜血管分割框架，能够克服视网膜血管分割中 Medical Image Computing and Computer-Assisted Intervention. 存在的细小血管提取不完整、分割不准确的问题 Munich,2015:234 从而辅助医生开展早期眼底病变筛查.提出的框 [13]Guo CL,Szemenyei M,Hu H,et al.Channel attention residual U- Net for retinal vessel segmentation [J/OL].arXiv preprint (2020- 架与主干网络结构无关，也可以灵活扩展到其他 10-20)[2021-6-10].https://arxiv..0rg/abs/2004.03702 与拓扑结构相关的分割任务 [14]He K M,Zhang X Y,Ren S Q,et al.Deep residual learning for image recognition /2016 IEEE Conference on Computer Vision 参考文献 and Pattern Recognition (CVPR).Las Vegas,2016:770 [1]Cong M,Wu T,Liu D,et al.Prostate MR/TRUS image [15]Vaswani A,Shazeer N,Parmar N,et al.Attention is all you need//85 84 83 82 81 1 2 3 4 5 6 7 8 9 1011121314151617181920 Singel task network +Skeleton extraction +Structure loss Training iteration (a) F1/ % 82.5 82.0 81.5 81.0 80.5 80.0 79.5 79.0 1 2 3 4 5 6 7 8 9 1011121314151617181920 Training iteration (b) F1/ % Singel task network +Skeleton extraction +Structure loss 图 5    消融实验中每轮训练后在不同验证集上的 F1 值. （a）DRIVE； （b）CHASEDB1 Fig.5    F1 on the validation set after each training iteration in ablation experiments: (a) DRIVE; (b) CHASEDB1 表 3 第二组消融实验结果 Table 3   Results of the second ablation experiments Network DRIVE CHASEDB1 F1/% Se/% Acc/% AUC/% F1/% Se/% Acc/% AUC/% ResNet34 83.1 83.7 97.1 98.8 82.0 84.5 97.1 99.1 ResNet18 82.9 83.5 97.0 98.7 81.7 83.8 97.6 99.0 VGG16 82.8 83.2 97.0 98.7 81.6 83.7 97.6 99.0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Epoch 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Epoch Loss 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0 Loss 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0 (a) (b) 图 6    框架在 DRIVE（a）和 CHASEDB1（b）数据集上的训练损失 Fig.6     Training  loss  of  the  framework  on  the  DRIVE  (a)  and CHASEDB1 (b) datasets 4    结论 本文提出了一种骨架图引导的多任务级联视 网膜血管分割框架，能够克服视网膜血管分割中 存在的细小血管提取不完整、分割不准确的问题. 从而辅助医生开展早期眼底病变筛查. 提出的框 架与主干网络结构无关，也可以灵活扩展到其他 与拓扑结构相关的分割任务. 参    考    文    献 [1] Cong  M,  Wu  T,  Liu  D,  et  al.  Prostate  MR/TRUS  image segmentation  and  registration  methods  based  on  supervised learning. Chin J Eng, 2020, 42（10）: 1362 （丛明, 吴童, 刘冬, 等. 基于监督学习的前列腺MR/TRUS图像分 割和配准方法. 工程科学学报, 2020, 42（10）：1362） Ma B Y, Jiang S F, Yin D, et al. Image segmentation metric and its application  in  the  analysis  of  microscopic  image. Chin J Eng, 2021, 43（1）: 137 （马博渊, 姜淑芳, 尹豆, 等. 图像分割评估方法在显微图像分析 中的应用. 工程科学学报, 2021, 43（1）：137） [2] Tso  M  O  M,  Jampol  L  M.  Pathophysiology  of  hypertensive retinopathy. Ophthalmology, 1982, 89（10）: 1132 [3] Yu  S,  Xiao  D,  Kanagasingam  Y.  Machine  learning  based automatic neovascularization detection on optic disc region. IEEE J Biomed Heal Inform, 2018, 22（3）: 886 [4] Becker  C,  Rigamonti  R,  Lepetit  V,  et  al.  Supervised  feature learning  for  curvilinear  structure  segmentation  // International Conference on Medical Image Computing and Computer-Assisted Intervention. Berlin, 2013: 526 [5] Tolias  Y  A,  Panas  S  M.  A  fuzzy  vessel  tracking  algorithm  for retinal  images  based  on  fuzzy  clustering. IEEE Trans Med Imaging, 1998, 17（2）: 263 [6] Soares  J  V  B,  Leandro  J  J  G,  Cesar  R  M,  et  al.  Retinal  vessel segmentation  using  the  2-D  Gabor  wavelet  and  supervised classification. IEEE Trans Med Imaging, 2006, 25（9）: 1214 [7] Sebbe  R,  Gosselin  B,  Coche  E,  et  al.  Segmentation  of  opacified thorax  vessels  using  model-driven  active  contour  //  2005 IEEE Engineering in Medicine and Biology 27th Annual Conference. Shanghai, 2006: 2535 [8] Pal  S,  Chatterjee  S,  Dey  D,  et  al.  Morphological  operations  with iterative  rotation  of  structuring  elements  for  segmentation  of retinal  vessel  structures. Multidimens Syst Signal Process,  2019, 30（1）: 373 [9] Chang C C, Lin C C, Pai P Y, et al. A novel retinal blood vessel segmentation  method  based  on  line  operator  and  edge  detector  // 2009 Fifth International Conference on Intelligent Information Hiding and Multimedia Signal Processing. Kyoto, 2009: 299 [10] Zhang Y S, Chung A C S. Deep supervision with additional labels for retinal vessel segmentation task // International Conference on Medical Image Computing and Computer-Assisted Intervention. Granada, 2018: 83 [11] Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. Munich, 2015: 234 [12] Guo C L, Szemenyei M, Hu H, et al. Channel attention residual U￾Net for retinal vessel segmentation [J/OL]. arXiv preprint (2020- 10-20) [2021-6-10]. https://arxiv.org/abs/2004.03702 [13] He K M, Zhang X Y, Ren S Q, et al. Deep residual learning for image  recognition  //  2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, 2016: 770 [14] [15] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need// · 1250 · 工程科学学报，第 43 卷，第 9 期