Abstract: In computer aided diagnosis (CAD) applications the goal is to detect structures of interest to physicians in medical images: e.g. to identify potentially malignant lesions in an image (mammography, lung CT, Colon CT, heart ultrasound, etc.). In an almost universal paradigm, this problem is addressed by a 5 stage system: 1. Segmentation to identify/extract the general area of interest; 2. Candidate generation which identifies suspicious unhealthy candidate regions of interest (ROI) from a medical image; 3. feature extraction that computes descriptive features for each candidate; 4. classification that differentiates candidates based on candidate feature vectors; 5. visual presentation of CAD findings to the radiologist in order for him to accept or reject the CAD findings. For the fourth stage, many standard algorithms (such as support vector machines (SVM), back-propagation neural nets, kernel Fisher discriminants) have been used to learn classifiers for detecting malignant structures. However, these general-purpose learning methods either make implicit assumptions that are commonly violated in CAD applications, or cannot effectively address the difficulties arisen when learning a CAD system. Non-IID Data Traditional learning methods almost universally assume that the training samples are independently drawn from an identical albeit unobservable underlying distribution (the IID assumption), which is often not the case in CAD systems. Due to spatial adjacency of the regions identified by a candidate generator, both the features and the class labels of several adjacent candidates are highly correlated. In this talk we present two recent proposed machine learning algorithms that successfully takes into account the correlation among candidates to significantly improve classification performance.
