Abstract
Chronic liver disease is a progressive disease, most of the time asymptomatic, and potentially fatal. In this chapter, an automatic procedure to stage the disease is proposed based on ultrasound (US) liver images, clinical and laboratorial data.
A new hierarchical classification and feature selection (FS) approach, inspired in the current diagnosis procedure used in the clinical practice, here called Clinical-Based Classifier (CBC), is described. The classification procedure follows the well-established strategy of liver disease differential diagnosis. The decisions are taken with different classifiers by using different features optimized to the particular task for which they were designed. It is shown that the CBC method outperforms the traditional one against all (OAA) method because it take into account the natural evolution of the hepatic disease. Different specific features are used to detect and classify different stages of the liver disease as it happens in the classical diagnosis performed by the medical doctors.
The proposed method uses multi-modal features, extracted from US images, laboratorial and clinical data, that are known to be more appropriated according to the disease stage we want to detect. Therefore, a battery of classifiers and features are optimized and used in a hierarchical approach in order to increase the accuracy of the classifier.
For the normal class we achieved 100% accuracy, for the chronic hepatitis 69.2%, for compensated cirrhosis 81.48%, and for decompensated cirrhosis 91.7%.