Data Sampling Imbalance with Steerable Wavelets for Abnormality Detection in Brain Images

A long standing goal within artificial intelligence application for medical imaging has been the ability for appropriate detecting abnormalities in MRI image of brains to support early diagnostics of cancer. This paper presents a solution relying on analysis of class imbalance in data sampling from brain image database instead of error statistics to improve accuracy of the abnormality detection. Here we use modification of training data set both for minority class and majority class to overcome under-segmentation and over-segmentation in detection of abnormality where abnormality is seen as minority class but its distribution is assumed unknown. In this approach, steerable wavelet based features are encoded with machine learning methods to allow the study of data sampling imbalance. In order to increase the detection sensitivity a set of wavelet features is selected from a number of feature sets in the learning task. The results with a benchmark medical image database show the effectiveness of the proposed method for abnormality detection in brain images.