Poster
Transcription
Poster
Oblique Random Forests for 3-D Vessel Detection Using Steerable Filters and Orthogonal Subspace Filtering Matthias Schneider1, Sven Hirsch1, Gábor Székely1, Bruno Weber2, and Bjoern H. Menze1 1 Computer Vision Laboratory, ETH Zurich, Switzerland 2 Institute of Pharmacology and Toxicology, University of Zurich, Switzerland COMPUTER AIDED AND IMAGE GUIDED MEDICAL INTERVENTIONS Motivation Vessel detection and segmentation Objectives • Crucial for many clinical diagnostic and planning tasks • Challenges: - Multiscale nature of blood vessels - Contrast inhomogeneities - Image noise and artifacts • Clinical studies on neurodegenerative diseases (e.g. Alzheimer’s disease) make use of high-resolution imaging to analyze microvascular structures of the brain. ► Efficient tools for vessel segmentation and analysis required - Vessel segmentation framework - Applicable for 2-D/3-D data - Efficient computation - Scalable (high-resolution datasets) Figure 1: Left: Medical images showing different vascular structures: Fundus photography (top) and fluoroscopy of the left coronary artery (bottom). Right: Scanning electron micrograph of a vascular corrosion cast from the macaque primary visual cortex [6]. Arteries are shaded in red and veins are blue. The circular image shows an axial slice of a cylindrical sample punched out of the cortex (superimposed cylinder) obtained by synchrotron radiation X-ray tomographic microscopy (srXTM) [9]. Method Machine Learning Framework Orthogonal Subspace Filters (OSF) • Learn “optimal” eigenfilters from local vessel patches • Orthogonal basis preserving maximum variance • Vessel appearance described in low-dimensional feature space Steerable Filter Templates (SFT) • Design parameterized filter templates similar to highly structured OSF eigenfilters in order to better model the problem and explain the data [5] • Parameterization: Gaussian derivatives up to order M Random Forest Classifier [1] • Orthogonal splits [1] - Optimal thresholds on randomly selected single features - Orthogonal 1-D hyperplanes • Advantages: - Explicit scale parameterization - Steerability allows for efficient directional filtering • Oblique splits [7] - Multidimensional hyperplanes - Linear regression with elastic-net penalty to learn multivariate (optimal) split direction [4] - Better generalization in areas with few samples Why Random Forests? • Efficient, scalable, and accurate • Continuous posterior (”confidence”) • Very few parameters involved • Estimates of feature importance and generalization error during training Figure 3: Machine learning framework for 3-D vessel segmentation using OSF and SFT features along with a random forest (RF) classifier. The steerable SFT filters can efficiently be applied along the normalized vessel direction (Rθ,φ). Figure 2: Visualization of 3-D filter templates. Top: OSF eigenfilters learned from vessel patches. Bottom: SFT templates at fixed scale up to order two. Templates are plotted for centered sagittal, coronal, and axial 2-D slices. Results σ =1 σ =2 σ =3 σ =4 σ =5 σ =6 σ =7 σ =8 10 10 −1 10−2 10−3 10−4 10−5 1 10 19 28 37 46 Feature index 55 64 k Gσ1,0,0 σk G1,1,0 k Gσ1,1,1 k Gσ2,0,0 k Gσ2,1,0 k Gσ2,1,1 σk G2,2,0 k Gσ2,2,1 k Gσ2,2,2 1 RF-OSF d=9 d = 27 d = 57 d = 102 RF-SFT M =1 M =2 M =3 M =4 Frangi Sato Otsu maximum F1 0.8 Precision 0 Importance Image Data • Four high-resolution 3-D datasets • Synchrotron radiation X-ray tomographic microscopy (srXTM) • Cylindrical samples of the murine somatosensory cortex • Dimensions (2048px)3, 700nm isotropic voxel spacing, ROI (256px)3 0.6 0.4 0.2 0 0.5 72 0.6 0.7 0.8 Recall Figure 4: Visualization of decision boundaries (solid black) for a two class problem (red/green) using a random forest classifier along with orthogonal splits (left), oblique splits in a single decision tree (center) and an ensemble of decision trees (right) [7]. The dashed black line indicates Bayes’ optimal border. Figure 6: Feature relevance (permutation importance [1]) of the RF-OSF (left) and RF-SFT model (right) on a logarithmic scale (oblique splits). Right: Precision-recall curves and optimal operating points w.r.t. F1 measure for RF-OSF and RF-SFT models (oblique splits) with varying number of features in comparison to (optimized) Frangi’s/Sato’s vesselness [2], and Otsu thresholding [8]. 0.9 0.95 1 Experiments Orthogonal Splits SFT Features • OSF model: Eigenfilters computed from 3000 randomly • Fast training • Second order Gaussian derivasampled vessel patches of size (19px)3 • Very fast split evaluation tives highly discriminative • SFT model: Gaussian derivatives up to order M=1,2,3,4 at • Relatively deep trees, particularly • Clearly outperforms OSF features three scales resulting in 9 (27, 57, 102) features for highly correlated features w.r.t. segmentation performance • RF training - Otsu thresholding used to compute training labels Oblique Splits - Training set: 4000 randomly sampled FG/BG samples • Split evaluation and finding optimal - 256 trees split parameters more expensive • Ground truth • Better generalization and accuracy - Semi-automatic active-contour segmentation tool • Smaller average tree depth - Manual corrections by expert on selected slices ► Oblique splits are superior and Figure 5: Comparison of oblique and - Used for evaluation and validation only! worth the additional computa- orthogonal split models w.r.t. average tree depth (left) and out-of-bag (OOB) tional effort during training 10 6 4 OSF SFT Feature class Oblique Random Forests • Efficient, scalable, and accurate predictor • Oblique splits favorable over univariate orthogonal splits - Better generalization in areas with few samples - Smaller average tree depth (faster prediction) • Superior to Hessian-based approaches 1 0 0 Steerable Filter Templates (SFT) • Parameterization of OSF eigenfilters using Gaussian derivatives • Preferable over OSF features for better problem modeling - Explicit scale parameterization (multiscale nature of vessels) - Efficient directional filtering (rotational invariant features) • Efficient feature extraction: separability + steerability 1.5 0.5 2 Machine learning framework for 3-D vessel segmentation Orthogonal Oblique 2 OOB error [%] Average tree depth 8 Conclusions 2.5 Orthogonal Oblique OSF SFT Feature class Figure 7: Visualization of the segmented cerebrovascular network for a single axial slice (top) and the entire 3-D test ROI (bottom) using different segmentation techniques. (a) Ground truth. (b) Frangi [2]. (c) RF-OSF (d = 102). (d) RF-SFT (M = 4). The segmentation results are rendered in error (right) for OSF and SFT features. 3-D (bottom) and outlined in red (top) along with the ground-truth contours in blue for three Errorbars indicate the standard deviation subregions within the axial slice (A-C). Red contours in (a) mark the Otsu labels [8] used for RF training. Black circles in the 3-D plots highlight prominent differences in the segmentation. over the four datasets. Acknowledgements This work has been funded by the Swiss National Centre of Competence in Research on Computer Aided and Image Guided Medical Interventions (NCCR Co-Me) funded by the Swiss National Science Foundation. References [1] Breiman, L.: Random forests. Machine Learning 45, 5–32 (2001) [2] Frangi, A., Niessen,W., Vincken, K., Viergever, M.: Multiscale vessel enhancement filtering. In: Wells, W., Colchester, A., Delp, S. (eds.) MICCAI’98. LNCS, vol. 1496, pp. 130–137. Springer, Berlin/Heidelberg (1998) [3] Freeman, W.T., Adelson, E.H.: The design and use of steerable filters. IEEE Trans Pattern Anal Mach Intell 13(9), 891–906 (Sep 1991) [4] Friedman, J.H., Hastie, T., Tibshirani, R.: Regularization paths for generalized linear models via coordinate descent. J Stat Softw 33(1), 1–22 (Feb 2010) [5] González, G., Fleurety, F., Fua, P.: Learning rotational features for filament detection. In: CVPR 2009. pp. 1582–1589 (Jun 2009) [6] Hirsch, S., Reichold, J., Schneider, M., Székely, G.,Weber, B.: Topology and hemodynamics of the cortical cerebrovascular system. J Cereb Blood Flow Metab (Apr 2012) [7] Menze, B.H., Kelm, B.M., Splittho, N., Koethe, U., Hamprecht, F.A.: On oblique random forests. In: ECML PKDD 2011. Springer (2011) [8] Otsu, N.: A threshold selection method from gray-level histograms. IEEE T Syst Man Cyb 9(1), 62–66 (Jan 1979) [9] Reichold, J., Stampanoni, M., Keller, A.L., Buck, A., Jenny, P., Weber, B.: Vascular graph model to simulate the cerebral blood flow in realistic vascular networks. J Cereb Blood Flow Metab 29(8), 1429–1443 (Aug 2009)