1 Fibrosis detection from ultrasound imaging. The influence of necro- inflammatory activity and steatosis over the detection rates. Cristian Vicas, Sergiu Nedevschi, Monica Lupsor, Radu Badea Technical University of Cluj-Napoca 3’rd Medical Clinic Cluj Napoca {cristian.vicas,sergiu.nedevschi}@cs.utcluj.ro {mmlupsor, rbadea2003}@yahoo.com Abstract Diagnosing liver fibrosis using non invasive procedures is challenging because the visual aspects in US imaging between healthy and fibrosis liver are very much alike. In this paper texture analysis (texture feature computation and texture classification) are employed in order to increase the diagnosis value of US examination. An overview on fibrosis detection is necessary in order to determine and evaluate the best approach. Biopsy and METAVIR score are used to assess the liver pathology. The influence of steatosis and necro-inflamatory activity over the fibrosis detection is also investigated. Four feature selection methods based on gain ratio, chi squared statistic, correlation and symmetrical uncertainty are evaluated. The results show that fibrosis, steatosis and activity alter the US image texture and implicitly the texture features. It seems that the best approach in liver fibrosis identification is to build imagistic models for each fibrosis grade. 1. Introduction The fibrosis is the scarring response formed in the chronic injury of any cause. It is a dynamic process, with a possibility of reversibility. For the moment, the golden standard in evaluating fibrosis is the liver biopsy. Using the liver biopsy one can establish with certainty the diagnosis, one can assess the severity of necroinflamation and fibrosis and one can distinguish the simultaneous liver diseases. On the other hand, it is an invasive procedure, with possible side-effects. [1],[2] An alternative examination procedure is ultrasound imaging. In fibrosis the visual aspects of healthy/affected liver are very similar so, the diagnosis value of the ultrasound imaging is relatively low in these diseases. Image processing techniques are used to study ultrasound images and to improve the diagnosis value of ultrasound in diffuse liver diseases. Features are computed over image textures and the images are classified based on the computed feature vectors. One fractal dimension based algorithm (local multifractal morphological exponents) is used for the first time in fibrosis level detection field. This paper is structured in the following manner: in chapter 2 are presented shortly the feature computation algorithms and the support vector machine classifier, in chapter 3 are presented some experimental results and chapter 4 contains discussions and conclusions. 2. Texture features and classification algorithm Eight feature computation algorithms are employed: histogram statistics, grey tone difference matrix, grey level co-occurrence matrix, multifractal differential box counting, morphological multifractal exponents, multi resolution fractal dimension, Law’s energy measures and wavelet transform. Most of these algorithms are common in texture analysis field so they will be presented briefly. 2.1. First and second order statistics Histogram statistics.[3] The shape of the gray level histogram can provide clues about the image. Central moments are derived from the histogram in order to characterize the texture. These are mean, variance, skewness, kurtosis, energy and entropy. Grey-Tone difference matrix [4]. A Grey-Tone Difference Matrix a column vector containing G elements where G is the total number of gray levels. Its entries are computed measuring the difference between the intensity level of a pixel and the average intensity computed over a square window centered at the pixel.