A novel image representation method for liver tumor classification

Computer aided diagnosis (CAD) has been important more than ever for accurate diagnosis of liver tumors. The paper presents a novel image representation method for classifying normal livers and livers with tumors. It starts by capturing region of interesting (ROI) for individual livers, on which patches are extracted densely. Histogram of oriented gradients (HOG) and intensity are then extracted as patch features. Taking the feature clustering centers in the training images as coding dictionary, sparse coding is used as a coding scheme for the patch extracted from both train and test images. And an effective image representation is then generated based on bag of features (BOF). In this study, an optimized coding method based on the dictionary elements nearby are utilized, which accelerate the coding procedure. The experimental results demonstrate that the proposed image representation method achieves higher classification rate.