VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 MEDICAL IMAGE ANALYSIS Final Project - 3D Breast Ultrasound Segmentation Students: Flávia Dias Casagrande Marcel Sheeny de Moraes
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Outline • Introduction • Active Contour • Level Set Segmentation • Algorithm Implemented • Preprocessing • Processing • Evaluation • Results • Conclusion
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Introduction • Objective: 3D breast cancer ultrasound segmentation • 5 images and their ground truths • Propose segmentation method
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Active Contour • Geometric alternative for “ snakes ” • Geometric flow (Partial Differential Equations) • Geodesic active contours Ref: http://scialert.net/fulltext/?doi=ajaps.2011.101.111&org=12
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Level Set • Moving fronts: curve propagation • Idea: represent the evolving contour using a signed function whose zero corresponds to the actual contour. Original front Level set function Ref: https://leticiateran.files.wordpress.com/2013/04/63a28-2.png
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Level Set • The zero level is spread, reflecting the propagation of the contour • For one direction: fast marching Evolution of the zero level contour (red Evolution of the level set function (the red curve) curve is the zero level contour) Ref: http://www.imagecomputing.org/~cmli/code/
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Algorithm Implemented • ITK: Pipeline of filters Input image Anisotropic Gradient Sigmoid Diffusion Magnitude Preprocessing Filter Filter Filter Binary Fast Marching Geodesic Active Threshold Level Set Filter Contours Filter Filter Seed point Output image
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Smoothing • itk::CurvatureAnisotropicDiffusionImageFilter • Noise reduction • Parameters: time step, number of iterations, conductance parameter Axial – Sagittal – Coronal Smoothing results for pacient 3
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Gradient Magnitude • itk::GradientMagnitudeRecursiveGaussianImageFilter • Boundaries enhancement • Parameters: standard deviation 𝜏 Axial – Sagittal – Coronal Gradient Magnitude results for pacient 3
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Sigmoid • itk::SigmoidImageFilter • Contrast enhancement • Parameters: 𝛽 and 𝛾 Axial – Sagittal – Coronal Sigmoid results for pacient 3
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Segmentation . ppt www.na-mic.org/Wiki/.../Insight- Fast Marching • itk::FastMarchingImageFilter Δ x • Estimate an initial rough contour • Parameters: seed point, speed constant and initial distance Axial – Sagittal – Coronal Fast marching results for pacient 3
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Geodesic Active Contour • itk::GeodesicActiveContourLevelSetImageFilter • Refine initial approximation of contour • Parameters: propagation scaling (inflation), curvature (smoothing) and the advection Axial – Sagittal – Coronal Geodesic active contours results for pacient 3
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Binary Threshold • Itk::BinaryThresholdImageFilter • Produce a binary mask • Parameters: lower and upper thresholds, outside and inside values Axial – Sagittal – Coronal Binary thresholds results for pacient 3
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Evaluation • itk::LabelOverlapMeasuresImageFilter • Dice coefficient • Jaccard index • Confusion matrix: specificity and sensitivity Measurement Image 1 Image 2 Image 3 Image 4 Image 5 Sensitivity 0.670722 0.838789 0.737402 0.215896 0.873426 Specificity 0.999932 0.999647 0.999894 0.999997 0.99976 Jaccard 0.520453 0.335186 0.545407 0.212369 0.452144 Dice 0.684603 0.502082 0.705842 0.350337 0.622726 • Sensitivity Average: 0.667247 • Specificity Average: 0.999846 • Jaccard Index Average: 0.4131118 • Dice Coefficient Average: 0.573118 • Computation time: around 145s (from reading to saving file)
Results Pacient 1 Pacient 2
Results Pacient 3 Pacient 4
Results Pacient 5
VIBOT9 - Universitat de Girona Medical Image - 3D Breast Ultrasound Segmentation May 13th, 2015 Conclusion • Impressions from the implemented segmentation method • Advantages : • Accuracy/stability when contour length changes • Deal with corners • Topology changes are automatically handled • Disadvantages : • Really sensitive to noise • Highly sensitive to the placement of seed points. • Impressions of the project
References • Geodesic Active Contours. Vicent Caselles, Ron Kimmel, and Guillermo Sapiro. Int. J. Comput. Vision .1997. • Cerebral Artery Segmentation with Level Set Methods . H. Ho, P. Bier, G. Sands, and P. Hunter. Proceedings of Image and Vision Computing New Zealand. 2007 • Fast geodesic active contours. Goldenberg, R. Dept. of Comput. Sci., Technion-Israel Inst. of Technol., Haifa, Israel. 2001. • Image Segmentation Using Active Contour Model and Level Set Method Applied to Detect Oil Spills . M. Airouche, L. Bentabet and M. Zelmat. Proceedings of the World Congress on Engineering. 2009. • Methods based on PDEs and wavelets in medical image segmentation . <https://leticiateran.wordpress.com/>. 2015. • ITK Image Segmentation <www.na-mic.org/Wiki/.../Insight- Segmentation.ppt> • Moving interfaces and boundaries: level set methods and fast marching methods. <https://math.berkeley.edu/~sethian/2006/level_set.html>
Recommend
More recommend
