A Roi-bag Approach for Automatic Liver Cirrhosis Diagnosis Using Ultrasound Images

In this paper, we present a soft-computing approach to improve the accuracy in recognizing the state of the liver based on a clinical ultrasound image. The detection of regions of interest (ROIs), which significantly reveal liver aberrance, remains a challenge since the image quality is relatively low in the real-world cases. Instead of using a single ROI, in this work, the liver area is divided into multiple ROIs to represent the image as a bag of ROIs. To cluster all training ROI texture features regardless of their liver states, an ROI background codebook is learned using a fuzzy clustering algorithm. The class-specific codebooks are then adapted from the background codebook in order to construct the ROI-Bag descriptor. To apply the genetic algorithm (GA), every training ROIs in a ROI-Bag is then re-sampled to construct multiple ROI-bags with trained individual weak classifiers. The ensemble of weak classifiers finally constitutes a strong two-class classifier. A cascade of these two-class classifiers is also structured to improve the performance of multi-class classification. According to the clinical gold standard, when diagnosing chronic liver diseases, experiments based on a set of labeled ultrasound images, which were verified by the liver biopsy, demonstrate the effectiveness of the proposed method.