Segment 2D and 3D Filaments by Learning Structured and Contextual Features

Segment 2D and 3D Filaments by
Learning Structured and Contextual Features



Lin Gu, Xiaowei Zhang, He Zhao, Li Cheng,
@Machine Learning For Bioimage Analysis Group, BII, A*STAR,Singapore
 and
Huiqi Li,
@Beijing Institute of Technology, China

  • 2D Segmentation Results

    • Datasets: DRIVE, STARE, and 2D Neurons.
    Competing methods:
    1. SF: our approach using only structured features.
    2. SF + context distance: our approach using both structured features and context distance features.
    3. Kernel Boost [1] that learns discriminative convolutional filters based on the Gradient Boosting framework.
    4. Optimally Oriented Flux (OOF) [2], an unsupervised method with manually designed filter that is widely used for delineating tubular structures.
    5. IUWT [3], an unsupervised method based on isotropic undecimated wavelet transform.
    6. Eigen [4], a multiscale Hessian-based method.
    7. T2T [5], a supervised method based on pixel classification, medial sub-tree generation, and global tree linking components.
    8. Structured Edge (SE) [6], a structured edge detection technique using random forests.
    9. B-COSFIRE [7], selectively responds to vessels by computing the weighted geometric mean of the outputs by applying a pool of Difference-of-Gaussian filters.
    10. LCMBoost [8], an iterative learning-based approach that boosts the performance of an existing base segmentation method.

    Performance Statistics:

    Table 1: Performance statistics of 2D segmentation using F1 measure (%), Precision (%), Recall (%), Specificity (%) and MCC. Performance statistics


    Visual Results:

    2D visual results Figure 1: Exemplar results on segmenting 2D retinal and neuronal images. (a): Input images; (b): Ground-truth; (c): Probability maps of our approach (SF + context distance variant); (d & e) Error images of our approach (SF + context distance & SF only); (f) Error images of Kernel Boost; (f) Error images of SE. Here green denotes false alarm and the magenta denotes the missing error.

  • 3D Segmentation Results

    • Dataset: BigNeuron Gold166.
    Competing methods:
    1. SF: our approach using only structured features.
    2. SF + context distance: our approach using both structured features and context distance features.
    3. Adaptive Ehancement [9], adaptive image enhancement method specifically designed to improve the signal-to-noise ratio of several types of 3D neuronal and vessel images.
    4. gray-weighted distance transform (GWDT) [10], a distance transform used in APP2 to enhance the input image.
    5. Regression Tubularity [11], which could be considered as an extension of Kernel Boost, formulates the linear structure centreline detection as a regression problem.

    Results using F1 measure (%):
    SF SF + context distance Adaptive Ehancement GWDT Regression Tubularity
    79.38 ± 9.75 79.89 ± 9.33 58.53 ± 14.27 75.77 ± 10.52 65.75 ± 12.48

    Visual Results:

    3D visual results Figure 2: Exemplar 3D neuronal segmentation results on Gold166 dataset. (a & f): Input images with ground-truth in blue; (b & g): Results of our SF + context distance variant; (c & h) Results of Adaptive Enhancement; (d & i) Results of GWDT; (e & j) Results of Regression Tubularity.


    Video Illustration:

    (b) SF + context distance

    (c) Adaptive Enhancement

    (d) GWDT

    (e) Regression Tubularity

    (g) SF + context distance

    (h) Adaptive Enhancement

    (i) GWDT

    (j) Regression Tubularity


    Exemplar segmentation results on 3D neuronal images from the BigNeuron Dataset.

  • References:

    • [1] C. Becker, R. Rigamonti, V. Lepetit, and P. Fua. Supervised feature learning for curvilinear structure segmentation. In MICCAI, 2013.
    [2] M. Law and A. Chung. Three dimensional curvilinear structure detection using optimally oriented flux. In ECCV, 2008.
    [3] P. Bankhead, C. Scholfield, J. McGeown, and T. Curtis. Fast retinal vessel detection and measurement using wavelets and edge location refinement. PLoS ONE, 2012.
    [4] A. Frangi, W. Niessen, K. Vincken, and M. Viergever. Multiscale vessel enhancement filtering. In MICCAI, 1998.
    [5] S. Basu, A. Aksel, B. Condron, and S. Acton. Tree2Tree: Neuron segmentation for generation of neuronal morphology. In ISBI, 2010.
    [6] P. Dollar and C. L. Zitnick. Fast edge detection using structured forests. In IEEE TPAMI, vol 37, no. 8, pp. 1558--1570, 2015.
    [7] G. Azzopardi, N. Strisciuglio, M. Vento, N. Petkov. Trainable COSFIRE filters for vessel delineation with application to retinal images. In Medical Image Analysis, vol 19, no. 1, pp. 46--57, 2015.
    [8] L. Gu and L. Cheng. Learning to boost filamentary structure segmentation. In ICCV, 2015.
    [9] Z. Zhou, S. Sorensen, H. Zeng, M. Hawrylycz, and H. Peng, Adaptive image enhancement for tracing 3D morphologies of neurons and brain vasculatures, Neuroinformatics, vol. 13, no. 2, pp. 153–166, 2015.
    [10] H. Xiao and H. Peng, APP2: automatic tracing of 3D neuron morphology based on hierarchical pruning of a gray-weighted image distance-tree, Bioinformatics, vol. 29, no. 11, pp. 1448–54, 2013.
    [11] A. Sironi, E. Turetken, V. Lepetit, and P. Fua, Multiscale centerline detection, In IEEE TPAMI, 2016.