A non-parametric method based on NBNN for automatic detection of liver lesion in CT images

Abstract

An automatic liver lesion detection method for CT images is presented, which need not learn the model parameters and segment liver region. The lesion detection problem is formulated as finding a region with maximal score. The developed method employs an over-segmentation algorithm to generate the superpixels (small regions) and adapts the Naive Bayes Nearest Neighbor (NBNN) classifier to score the superpixels. Then, the connected superpixels with positive scores are aggregated as the detected regions. The performance of the method is evaluated on a data set consisting of 442 CT slices of 129 patients acquired in portal venous phase of contrast enhancement. The pixel-wise accuracy for classification and recall for detection can achieve 93% and 62%, respectively. The method can work well for hyperdense, hypodense, and heterogeneous liver lesions.