Feature Selection Methods in Image-Based Screening for the Detection of Hashimoto’s Thyroiditis in First-Contact Hospitals

Abstract

In this paper, results of dimension reduction in feature space for thyroid ultrasound images using the heuristic identification of noisy variables, testing the significance of correlation coefficients and the method of Hellwig index of information capacity, have been compared. The best results were achieved using the Hellwig method. It enabled us to choose only 3 features from a large set of 283 discriminant ones. Classifiers built on the basis of this reduced set of features have the highest classification sensitivity (0,82 ) and the highest classification specificity (0,83 ) in comparison to other reduced datasets that we used in our research. Results showed that the Hellwig method can be used as an effective process for dimension reduction in feature space in classification of thyroid ultrasound images.