Thyroid Nodule Classification Using Steerable Pyramid–Based Features From Ultrasound Images

Abstract

The most common endocrine cancer is thyroid cancer. The incidental rate of thyroid cancer has significantly increased during the past few decades. Timely identification and suitable treatment are essential for better outcome. High-resolution ultrasound is the preferred modality for the detection of thyroid nodules as it has the capability of locating tiny nodules. This article proposes a feature extraction method by integrating steerable pyramid decomposition and cooccurrence matrix features for the characterization of the thyroid nodule. Steerable pyramid decomposition is carried out both in time domain and frequency domain. Textural features are obtained from the pyramid at different levels and with different filters. ReliefF method is used for feature selection. Support vector machine is used to classify the thyroid nodule as benign or malignant, and its performance is evaluated using accuracy, sensitivity, specificity, positive predictive value, negative predictive value, false alarm rate, and F1_score. The proposed approaches are tested on a dataset containing 110 thyroid ultrasound images (benign, malignant, and borderline cases). A very high overall accuracy of 99.08% with 100% sensitivity (malignant nodule detected as malignant) and 98.16% specificity (benign nodule detected as benign) is obtained for features extracted from steerable pyramid coefficients through convolution using sp1 filter at level 3. Experimental results clearly indicate that steerable pyramid–based cooccurrence matrix features can effectively describe the distinctive nature of the thyroid nodule in ultrasound image.