In this talk I would like to address two problems in the context of developing and applying new signal processing techniques in biomedical research: 1. Efficient feature extraction using new steerable filters; 2. Reducing feature variation for better cancer-related biomarker discovery from mass spectrometry data. In the first part, I will focus on extracting local orientation information from images. This is still a difficult task even when we apply the newest idea of steerable filter. By attacking this problem from the standpoint of the sampling theory, we propose a novel set of filters based on Gaussian functions both in 2-D and 3-D space. The new filters achieve the lower bound of the uncertainty principle and provide higher orientation resolution yet with lower computational complexity than current steerable filters. Comparisons with related approaches and different applications such as facial analysis are shown.

The second part of the talk is about building mutual correspondence among features from multiple mass spectrometry data sets for the purpose of cancer-related biomarker discovery. The key point is to find a standard template from multiple feature sets (which are influenced by noise as well as biological and experimental variations) for matching purpose. After reviewing current methods, I will introduce a new approach based on the concept of diffusion and scale space representation. Then, I will validate our new approach using two different methods and compare our method with current methods. I will finish the talk by discussing some research topics planned for the future.