Representation and modelling of cirrhosis and hepatocarcinoma from ultrasound images using texture-based methods
Aims: Non-invasive, image based detection of diseases is one of the most important issues in nowadays research. The purpose of this paper is to establish imagistic models for cirrhosis and hepatocarcinoma. This task will consist in specifying those textural parameters that are relevant for each class, respectivelly in analysing these parameters, for finding the most probable intervals of variation and statistical values like their average, their minimum and maximum value and standard deviation.
Methods: For features extraction, we use the Gray Level Cooccurence Matrix, edge-based statistics, the Hurst fractal index and the Wavelet transform. Bayesian Networks are used for establishing the relevant parameters and the most probable intervals. Also, other classification methods like Decision Trees and Multilayer Perceptron are applied for evaluating the automatic diagnosis accuracy. The experiments were done on sets of 100 ultrasound images per class, acquired using the same settings of the echograph, under a pre-established protocol.
Results: Applying Bayesian Networks, the recognition rate for HCC was 94.91%. The same method revealed the most relevant parameters for the decision: the GLCM second order parameters and the entropy of the wavelet sub-images. Applying Decision Trees, the recognition rate was 96.61% and with Multilayer Perceptron the accuracy was 98.92%. In the case of cirrhosis differentiation, the following GLCM parameters are relevant: homogeneity, entropy, contrast and energy.
Conclusions: The high influence of entropy, energy and local homogeneity demonstrate the inhomogeneous, complex structure in grey levels for tumour areas and cirrhotic liver parenchima, in accordance with the complexity of the tissue.