Purpose: Clinical thyroid nodules classification is based on the thyroid imaging reporting and data system (TI-RADS). Among the classes of thyroid nodules, class-4 thyroid nodule is of great uncertainty with malignant risk ranging from 5% to 80%. Class-4 thyroid nodules have three sub-classes, namely 4a, 4b and 4c with different risk of malignancy. Accurate classification of TI-RADS class-4 thyroid nodules through machine learning approach is cost effective and is of clinical importance.
Methods: In this abstract, a k-nearest neighbor (KNN) model-based classifier was applied for classification of class 4 thyroid nodules. Patients with class-4 thyroid nodules were collected and classified into TI-RADS 4a, 4b and 4c according to ultrasound-guided fine-needle aspiration biopsy results. Ultrasound images for the enrolled patients were obtained. Pre-processing steps were as follows: a 128*128-pixel image containing the nodule was extracted from the original ultrasound image in order to remove annotations and markers. Artifacts were removed and data normalization was performed. A ternary label with -1 for TIRADS 4a, 0 for TI-RADS 4b and 1 for TI-RADS 4c was used. After pre-processing, image data were divided into training set and testing set. A KNN model was configured with euclidean distance for distance measure. Cross-validation was applied to find the optimal value of k for the model in the training set. Then, the KNN model with optimal value of k was trained in the training set. Performance evaluation of the fine-tuned KNN model-based classifier was tested in the testing set.
Results: The results demonstrated that the KNN model-based classifier could accurately and effectively classify TI-RADS class-4 thyroid nodules with an averaged accuracy of 85%. And the accuracies for each of the sub-class was 83.3%, 88.8% and 80% respectively.
Conclusion: The proposed KNN model-based classifier can be used for classification of the thyroid nodules clinically with fine-tuned model parameters.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by National Key Research and Development Program of China (2016YFC0103400), Key Research and Development Program of Shandong Province (2017GGX201010), Jianfeng Q. was supported by the Taishan Scholars Program of Shandong Province (TS201712065).