Classification of Ultrasound Thyroid Nodule Images into Benign and Malignant Lesions

Thyroid cancer is generally diagnosed by visual inspection of ultrasound images. The proposed method is to automate the classification of thyroid nodules into cancerous or not. This is to assist radiologists. Value of Peak Signal to Noise Ratio (PSNR) for noise filters were compared. The Gaussian filter had the highest PSNR value and was applied to segmented images. Features of Elliptical fit, histogram, morphological, Gray Level Run Length Matrix (GLRLM) and Gray Level Co-occurrence Matrix (GLCM) are extracted from the pre-processed ultrasound images. F-score and Sequential Forward Feature Selection (SFFS) methods were used for feature selection. The optimal Cost function (C) and gamma parameters were obtained using the Grid Search algorithm. These are then used for the training the classifiers. Support Vector Machine (SVM) classifier and Linear Discriminant Analysis (LDA) performance was noted for both the feature selection methods and compared. An accuracy of 0.98 and 1.0 was obtained using LDA and SVM respectively with SFFS. An accuracy of 0.95 was obtained for both LDA and SVM with F-score.