IBM Almaden Research Center  Fei Wang's Homepage 
AALIM: Multimodal Mining for Cardiac Decision Support  
Diagnostic decision support is still very much an art for physicians in their practices today due to lack of quantitative tools. AALIM is a decision support system for cardiology that exploits the consensus opinions of other physicians who have looked at similar patients, to present statistical reports summarizing possible diagnoses. The key idea behind our statistical decision support system is the search for similar patients based on the underlying multimodal data. In this paper, we describe the AALIM decision support system and the underlying multimodal similarity search used for cardiac data sets. [webpage]  
[Wang et al. ICDAR 09] [Kumar et al. CVPR 09]  
Divergence Measures for Groupwise Shape Alignment  
The work fits into the general class of approaches that avoid explicit point correspondences for nonrigid alignment through the use of divergence measures between probability distributions formed around point sets. Specifically, we propose several divergence measures, such as JensenShannon (JS), JensenRenyi (JR) and the CDFbased JS divergence to measure the cohesiveness between the density functions to obtain the nonrigid deformation. The densitybased approaches are relatively more robust to the shapes of different sizes and to the presence of missing features. Furthermore, if an unbiased information theoretic measure is chosen to quantify the multiple densities representing the shapes, the matching results can potentially be unbiased to any of the given pointsets.  
[Wang et al. IPMI 09] [Wang et al. MICCAI 09]  
Biomedical Image Registration and Segmentation  
Much of this research has been concerned with the investigation of information
theoretic image registration methods, which employ entropic measures to quantify
the quality of alignment. The main issues of concern with these Shannon entropy
based measures are that they requires estimation of density distribution, which
in general may or may not exist, and the estimates however converge to
the true density only under some conditions; The results using information
measures defined based on Shannon entropy not only are sensitive to the noise
but also dependent on the field of views (FOVs). For those reasons, we develop a novel measure of information in a random variable based on its cumulative distribution that we dub cumulative residual entropy (CRE). This measure parallels the wellknown Shannon entropy but has the following advantages: (1) it is more general than the Shannon Entropy as its definition is valid in the discrete and continuous domains, (2) it possess more general mathematical properties and (3) it can be easily computed from sample data and these computations asymptotically converge to the true values. Based on CRE, we define the crossCRE (CCRE) between two random variables, and apply it to solve the image alignment problem for parameterized (3D rigid and affine) transformations. The key strengths of the CCRE over using the mutual information (based on Shannon's entropy) are that the former has significantly larger tolerance to noise and a much larger convergence range over the field of parameterized transformations. 

[Wang et al. IJCV 07] [Wang et al. MICCAI 05]  
Volumetric Visualization using Cylindrical Harmonics  
Realtime generation of Digitally Reconstructed Radiographs (DRRs) is crucial in intraoperative applications requiring matching of preoperative 3D data to 2D Xray images acquired intraoperatively. As part of the 3D2D medical image registration project, we presented a very fast algorithm for generating (DRRs) compared with conventional raycasting or the voxel projection technique. Our algorithm involves representing the preoperative 3D data set in a cylindrical harmonic representation and then projecting each of these harmonics from the chosen projection point to construct a set of 2D projections whose superposition is the DRR of the data set in its reference orientation. The key advantage of our algorithm over existing algorithms is that in our method, once the projection set is generated from an arbitrarily chosen point of projection, DRRs of the underlying object at arbitrary rotations are simply obtained via a complete exponentially weighted superposition of the set. This leads to tremendous computational savings over and above the basic computational advantages of the algorithm involving the use of truncated cylindrical harmonic representation of the data.  
[Wang et al. MICCAI 02]  