Non-invasive automated 3D thyroid lesion classification in ultrasound: a class of ThyroScan™ systems

Ultrasound-based thyroid nodule characterization into benign and malignant types is limited by subjective interpretations. This paper presents a Computer Aided Diagnostic (CAD) technique that would present more objective and accurate classification and further would offer the physician a valuable second opinion. In this paradigm, we first extracted the features that quantify the local changes in the texture characteristics of the ultrasound off-line training images from both benign and malignant nodules. These features include: Fractal Dimension (FD), Local Binary Pattern (LBP), Fourier Spectrum Descriptor (FS), and Laws Texture Energy (LTE). The resulting feature vectors were used to build seven different classifiers: Support Vector Machine (SVM), Decision Tree (DT), Sugeno Fuzzy, Gaussian Mixture Model (GMM), K-Nearest Neighbor (KNN), Radial Basis Probabilistic Neural Network (RBPNN), and Naive Bayes Classifier (NBC). Subsequently, the feature vector-classifier combination that results in the maximum classification accuracy was used to predict the class of a new on-line test thyroid ultrasound image. Two data sets with 3D Contrast-Enhanced Ultrasound (CEUS) and 3D High Resolution Ultrasound (HRUS) images of 20 nodules (10 benign and 10 malignant) were used. Fine needle aspiration biopsy and histology results were used to confirm malignancy. Our results show that a combination of texture features coupled with SVM or Fuzzy classifiers resulted in 100% accuracy for the HRUS dataset, while GMM classifier resulted in 98.1% accuracy for the CEUS dataset. Finally, for each dataset, we have proposed a novel integrated index called Thyroid Malignancy Index (TMI) using the combination of FD, LBP, LTE texture features, to diagnose benign or malignant nodules. This index can help clinicians to make a more objective differentiation of benign/malignant thyroid lesions. We have compared and benchmarked the system with existing methods.