Seminars

Multi-view medical imaging typically refers to imaging modalities where x-ray projection images of the organ(s) of interest are acquired from two or more viewing angles. The need to provide the clinician with accurate quantitative information from noise-prone low dimensional projection data gives rise to many interesting computational problems. In this talk, I will focus on two related problems from this area. A key problem in multi-view projection imaging is to find correspondences between 2D image points in multiple views. I will introduce this by focusing on Brachytherapy Seed Localization problem in prostate cancer treatment. Determination of correspondences between projections of "seeds" in multiple views given only approximately accurate orientation geometry information enables seed implant reconstruction and quantification of dose deviations. I will first highlight the key challenges and then illustrate how these can be addressed via an effective computational modeling in a linear programming framework. Next, I will discuss a problem that is conceptually the inverse of the first one - that of determining the orientation geometry given known correspondences. Here, we want to accurately reconstruct coronary vessels given two angiographic views of the vessels with known corresponding vessel bifurcation points in both views. I will describe an efficient algorithm for this problem and illustrate how our modeling not only enables predicting expected reconstruction errors but also establishes a link between most existing algorithms. I will conclude my talk with a brief overview of future research directions.

Bio: Vikas is a PhD student in Computer Science and Engineering at SUNY-Buffalo. He is also a member of Toshiba Stroke Research Center, SUNY-Buffalo. His research interests are in algorithm design for medical and biological image analysis. (personal homepage: http://www.cse.buffalo.edu/~vsingh).