Interactive Segmentation For Image Guided Therapy Ohad Shitrit & Tsachi Hershkovich Superviser: Dr. Tammy Riklin Raviv Ben-Gurion University of the Negev
What are we going to speak about? Tsachi H. & Ohad S. Computed Tomography Motivation Mathematical introduction Problem definition Energy Gradient Descent The system Results Conclusion & Future work
Tsachi H. & Ohad S. Computed Tomography (CT) Spiral Cone-Beam Scanning for Computed Tomography. Ge Wang, 2003 (All rights reserved)
Tsachi H. & Ohad S. Computed Tomography (CT) Spiral Cone-Beam Scanning for Computed Tomography. Ge Wang, 2003 (All rights reserved) X-Ray Projection using radon transform
Tsachi H. & Ohad S. Computed Tomography (CT) Radon transform as one dimensional Fourier transform Reconstructing the image with the inverse Fourier transform
Why is there any need for interactive segmentation ? Tsachi H. & Ohad S.
Why is there any need for interactive segmentation? Volume estimation is critical for further treatment Therapist knowledge is essential for final decisions Fast and accurate analysis might save life Tsachi H. & Ohad S.
“Active Contour” But first things first… Tsachi H. & Ohad S.
Probabilistic Model Based on Gaussian Mixture(GM) Tsachi H. & Ohad S. We will define the Probability of a Voxel (3D pixel) to belong to the Object Or to the Background:
A CT scan Histogram of a brain with Cerebral hemorrhage Tsachi H. & Ohad S.
Level Set Function -
Tsachi H. & Ohad S. Mathematical Issues
Tsachi H. & Ohad S. [Riklin Raviv, Van Leemput, Menze, Wells, Golland, Medical Image Analysis, 2011]
Tsachi H. & Ohad S. To achieve the optimized segmentation we maximize the joint distribution: [Riklin Raviv, Van Leemput, Menze, Wells, Golland, Medical Image Analysis, 2011]
Tsachi H. & Ohad S. Energy Functional Using the following relationship : Allows us to formulate our problem as an energy minimization problem Summing all contributions from each voxel
Tsachi H. & Ohad S. Image Likelihood Term Assuming Gaussian Mixtures Model (GMM)
Tsachi H. & Ohad S. Smoothness Term Objects in nature are continuous Trade off between smoothness and sensitivity Fine tuning is needed
Tsachi H. & Ohad S. User Interaction Term
Tsachi H. & Ohad S. Gradient Descent The gradient descent is an iterative process which leads to the minimum of the Energy term.
Tsachi H. & Ohad S. Block Diagram – Entire System [Riklin Raviv, Van Leemput, Menze, Wells, Golland, Medical Image Analysis, 2011]
Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Entire System 3D
Tsachi H. & Ohad S. Performance DiceSensitivitySpecificityAccuracy Automatic 0.874± ± ± ± With user interaction 0.905± ± ± ±0.0022
Tsachi H. & Ohad S. Data Provided by Dr. Ilan Shelef, Department of Radiology, Soroka University Medical Center and Zlotowski Center for Neuroscience, Ben- Gurion University of the Negev. Modality: CT Brilliance 64 Resolution: 512x512xZ (Z = ) X x Y x Z = 0.48 x 0.48 x 3 [mm] Z axis with 1.5[mm] overlap
Conclusion Semi-automatic segmentation tool Probabilistic model User Interface Collaboration with Soroka Medical Center Tsachi H. & Ohad S.
Future work Adjustments to other modalities (MRI) Handle with various of structures User-Machine dialog in the medical world Tsachi H. & Ohad S.
User Interface
Questions? Thank you. Demo -