An optimization framework for classifier learning from image data for computer-assisted diagnosis

Abstract

In computer-assisted medical diagnosis it is often hard or even impossible to obtain a valid set of rules for disease classification by classical knowledge engineering methods. Alternatively, machine learning methods are applied to obtain classifiers from sets of data pre-classified by medical experts. Typically in a medical context, available data sets are imbalanced with respect to the possible classifications. E.g., in dermatology, there are only few data representing cases of malign melanoma vs. many cases representing benign nevi. Furthermore, there are different missclassification costs assigned to different classes. E.g., it is much more critical (i.e. costly) to erroneously classify a malign melanoma as benign than the other way around. We propose a universally applicable optimization framework that successfully corrects the error-based inductive bias of classifier learning methods on image data. The framework integrates several techniques of common optimization techniques, such as modifying the optimization procedure for inducer-specific parameters, modifying input data by an arcing algorithm, combining classifiers of several classifier learning methods (kNN, SVM and C4.5) with different settings according to locally-adaptive, cost-sensitive voting schemes. The framework is designed to make the learning process cost-sensitive and enforcing more balanced missclassification costs between classes. The framework was evaluated on image data for Barrett’s esophagus with promising results compared to the base learners.