Medical Imaging Informatics


Modeling: Sparse Deformable Models for Cardiac Motion Analysis

We introduce a new family of deformable models, using Laplacian coordinates as the internal force and sparsity constraints as the external force. Benefited from the sparsity techniques, these deformable models are able to handle outliers or gross errors during the deformation. This is an extension of my thesis work on sparse shape representation. It is applied to the analysis of cardiac motion, using tagged Magnetic Resonance Imaging.

  • Sparse Deformable Models with Application to Cardiac Motion Analysis. [PDF]
    Yang Yu, Shaoting Zhang*, Junzhou Huang, Dimitris Metaxas, Leon Axel.
    IPMI, 2013. * Corresponding author

Segmentation: Robust Shape Prior via Sparse Shape Composition

This project is a collaboration with Siemens HealthCare (Malvern, PA). We propose a novel shape prior method to do shape inference and refinement, based on sparse representations. The method can robustly handle outliers, model complex shape variations, and preserve shape details. It is successfully applied to locate 2D lung in X-ray and segment 3D liver in CT.

  • Deformable Segmentation via Sparse Shape Representation. [PDF] [Journal version] [Project] [Python code for shape prior modeling]
    Shaoting Zhang, Yiqiang Zhan, Maneesh Dewan, Junzhou Huang, Dimitris Metaxas and Xiang Zhou. MICCAI, 2011. MICCAI Young Scientist Award Finalist. Top 25 hottest articles in Medical Image Analysis in 2012 full year (journal version)

Simulation: 4D Cardiac Reconstruction and Blood Flow Simulation

We use high resolution CT data to reconstruct 4D motion of the left ventricle endocardial surface. This reconstruction framework captures the motion of the complex anatomical features, such as papillary muscles and the ventricular trabeculae, which allow us to quantitatively investigate their possible functional significance in clinical. We also simulate blood flow using these meshes ([normal case] and [abnormal case]).


Segmentation: Modeling Region Statistics for Robust Segmentation

This project is a collaboration with IDEA lab (Lehigh University, PA) and Brookhaven National Lab (Upton, NY). A 3D deformable model is proposed to effectively segment 3D medical images. Deformations of the model are derived from a linear system that encodes external forces from the Region of Interest (ROI). We also use Laplacian driven model and hierarchical shape prior information to improve its performance. (Please refer to Tian Shen's homepage for more information)


Compressed sensing: Optimization of Large Scale Inverse Problems

To minimize a smooth convex function regularized by the mixture of prior models, we decompose it into multiple simpler subproblems. Then these subproblems are efficiently solved by existing techniques in parallel. The result of the original problem is obtained from the weighted average of solutions of subproblems in an iterative framework. This algorithm is applied to efficiently solve the MR imaging reconstruction

  • Efficient MR Image Reconstruction for Compressed MR Imaging. [PDF] [Project] [Code, Matlab]
    Junzhou Huang, Shaoting Zhang and Dimitris Metaxas.
    MICCAI, 2010. MICCAI Young Scientist Award, oral presentation

Computer Vision and Graphics


Nonverbal Communication Computing: Robust Face Tracking and ASL

Understanding how people exploit nonverbal aspects of their communication to coordinate their activities and social relationships is a fundamental scientific challenge. To analyze such events automatically, we have developed a real-time and pose-free 3D face tracker, and use it for fatigue detection, American Sign Language recognition, emotion analysis, etc.

  • A Review of Motion Analysis Methods for Human Nonverbal Communication Computing. [PDF], Dimitris Metaxas, Shaoting Zhang. IVC, 2013.
  • Pose-free Facial Landmark Fitting via Optimized Part Mixtures and Deformable Shape Model. [PDF], Xiang Yu, Junzhou Huang, Shaoting Zhang, Wang Yan, Dimitris Metaxas. ICCV, 2013.

Visual Search: Query Specific Fusion for Large-Scale Retrieval

This project is a collaboration with NEC Lab America (Department of Media Analytics) (Cupertino, CA). We focus on the large-scale image retrieval problem. Traditional methods either use vocabulary tree for local features (e.g., SIFT) or hashing code for holistic features (e.g., GIST). It is hard to combine them because the feature characteristics and the algorithmic procedures are dramatically different. We propose a graph-based query-specific fusion approach, which is unsupervised and has few parameters. We have achieved state-of-the-art performance on several public datasets.

  • Query Specific Fusion for Image Retrieval. [PDF] [Project] [Code] [Online demo]
    Shaoting Zhang, Ming Yang, Timothee Cour, Kai Yu, Dimitris Metaxas.
    ECCV, 2012.

Visual Search: Image Annotation and Retrieval using Group Sparsity

We proposed a regularization based feature selection algorithm to leverage both the sparsity and clustering properties of features, and incorporate it into the image annotation task. The algorithm was tested on Corel5K and IAPR TC12. Some Matlab code is available here: [annotation code].

  • Automatic Image Annotation Using Group Sparsity. [PDF] [Slides] [Code]
    Shaoting Zhang, Junzhou Huang, Yuchi Huang, Yang Yu, Hongsheng Li and Dimitris Metaxas.
    CVPR, 2010. Oral presentation.

Deformable Models: Mesh Editing and Geometry Processing

This course project extended Prof. Sorkine's As-Rigid-As-Possible Surface Modeling by considering skeleton information. ARAP surface modeling can recover rotations and preserve edge lengths. In our method, the volume is also roughly kept by leveraging the skeleton information. I wrote this demo (10 KLOC) with C++, OpenGL, Qt, OpenMesh, Newmat and TAUCS: [code], [video]. We also developed multiresolution technique to accelerate the deformation process.

  • Skeleton Based As-Rigid-As-Possible Volume Modeling. [PDF] [Slides] [Code]
    Shaoting Zhang, Andrew Nealen and Dimitris Metaxas.
    EG, 2010. (short paper, oral presentation)

Deformable Models: vtkModeling, A Deformation Toolkit via VTK

vtkModeling is based on Visualization ToolKit 5.0. Basically it is a collection of deformation algorithms, like Laplacian Surface Editing, Moving Least Square, Mass Spring System, Meshless Model, etc. I implemented them by extending a VTK class (vtkAlgorithm). Thus all classes can be easily used as VTK filters. The source code is avaiable at sourceforge: [vtkModeling_code], and a technical report: [How to extend VTK]. We used these algorithms to reconstruct the surface of sparse tMRI and initialize a meshless deformable model: [meshless deformation].

  • LV Surface Reconstruction From Sparse tMRI Using Laplacian Surface Deformation and Optimization. [PDF] [Code_report] [Code]
    Shaoting Zhang, Xiaoxu Wang, Dimitris Metaxas, Ting Chen and Leon Axel.
    ISBI, 2009.

Deformable Models: Surgical Simulation and Virtual Reality

We developed a virtual reality system to simulate the surgery in the endoscopy environment, which includes deformable models, collision detection and response, cutting and force feedback technique. I was the team leader of this project, and mainly focused on deformable models and collision response. These algorithms were implemented in C++ and wxPython. PHANTOM Desktop was used as the haptic device. A compact version is available in sourceforge (deformable model, collision detection and response, and simulator): [Simulator_code].

  • Real-Time Simulation of Deformable Soft Tissue Based on Mass-Spring and Medial Representation.
    Shaoting Zhang, Lixu Gu, Pengfei Huang, Jianfeng Xu.
    CVBIA (ICCV Workshop), 2005.