Design of a Fuzzy Inference Based Ultrasound Image Analysis System for Differential Diagnosis of Thyroid Nodules

Abstract

This paper presents a Fuzzy Inference based Ultrasound Image Analysis System for differential diagnosis of Thyroid Nodules (TNs). Thyroid Ultrasound (TUS) images containing TNs are preprocessed to remove speckle noise and are enhanced using histogram equalization method. Nodule boundaries are identified using the canny edge detection technique and required Region of Interest is obtained using Adaptive Regularized Kernel Fuzzy C-means (ARKFCM) segmentation algorithm. Nineteen texture features are extracted from the segmented images. Best First (BF), Genetic Search (GS) and Greedy Step Wise (GSW) search methods are applied to select best subset of features. Selected features are fuzzified. A novel, fuzzy system is built to discriminate benign from malignant TNs by employing Mamdani model to draw inferences and centroid scheme for defuzzification. Class Based Association (CBA) concept is used to generate fuzzy inference rules. The developed multiple input, single output FIUIAS resulted in an accuracy of 98%.