A Novel System for Multi-level Crohn's Disease Classification and Grading Based on a Multiclass Support Vector Machine

Crohn's disease (CD) is a chronic inflammatory condition of the gastrointestinal tract that can highly alter patient's quality of life. Diagnostic imaging, such as Enterography Magnetic Resonance Imaging (E-MRI), provides crucial information for CD activity assessment. Automatic learning methods play a fundamental role in the classification of CD and allow to avoid the long and expensive manual classification process by radiologists. This paper presents a novel classification method that uses a multiclass Support Vector Machine (SVM) based on a Radial Basis Function (RBF) kernel for the grading of CD inflammatory activity. To validate the system, we have used a dataset composed of 800 E-MRI examinations of 800 patients from the University of Palermo Policlinico Hospital. For each E-MRI image, a team of radiologists has extracted 20 features associated with CD, calculated a disease activity index and classified patients into three classes (no activity, mild activity and severe activity). The 20 features have been used as the input variables to the SVM classifier, while the activity index has been adopted as the response variable. Different feature reduction techniques have been applied to improve the classifier performance, while a Bayesian optimization technique has been used to find the optimal hyperparameters of the RBF kernel. K-fold cross-validation has been used to enhance the evaluation reliability. The proposed SVM classifier achieved a better performance when compared with other standard classification methods. Experimental results show an accuracy index of 91.45% with an error of 8.55% that outperform the operator-based reference values reported in literature.