Computational grading of hepatocellular carcinoma using multifractal feature description.

Cancer grading has become an important topic in the field of image interpretation-based computer aided diagnosis systems. This paper proposes a novel feature descriptor to observe the characteristics of histopathological textures in a discriminative manner. The proposed feature descriptor utilizes fractal geometric analysis with four multifractal measures to construct an eight dimensional feature space. The proposed method employed a bag-of-feature-based classification model to discriminate a set of hepatocellular carcinoma images into five categories according to Edmondson and Steiner's grading system. Three feature selection methods were utilized to obtain the most discriminative features of codeword dictionary (codebook). Furthermore, we incorporated four other textural feature descriptors: Gabor-filters, LM-filters, local binary patterns, and Haralick, to obtain a benchmark of the accuracy of the classification. Two experiments were performed: (i) classifying non-neoplastic tissues and tumors and (ii) grading the hepatocellular carcinoma images into five classes. Experimental results indicated the significance of the multifractal features for describing the histopathological image texture because it outperformed other four feature descriptors. We graded a given ROI image by defining a threshold-based majority-voting rule and obtained an average correct classification rate around 95% for five classes classification.