An Artificial Intelligence Ultrasound Platform for Screening and Staging Thyroid C

Abstract

<h3>Purpose/Objective(s)</h3> Thyroid cancer is one of the most rapidly increasing cancer in the US, largely due to increased detection. BRAF mutation (V600E) is common in thyroid cancer and is a druggable mutation. The purpose of this study was to establish a multimodal artificial intelligence (AI) ultrasound platform that consists of radiomics, topological data analysis (TDA), machine learning (ML)TI-RADS features, and deep learning (DL) to predict malignancy, pathological, and genomic outcome in patients with thyroid cancer. <h3>Materials/Methods</h3> Between 2010 and 2021, 1,346 thyroid nodule images through routine diagnostic ultrasound from 784 patients. The ultrasound images were divided into two datasets, one for internal training and validation, and one for external validation. Malignancy was confirmed with samples obtained from fine needle biopsy. Pathological staging and mutational status were confirmed with operative reports and genomic sequencing, respectively. To achieve maximum malignancy prediction, a quad-model AI method was performed – (1) radiomics which employs high-throughput quantitative analysis of image features, (2) TDA which analyzes geometric relationships between data points in images, (3) machine learning (ML) which algorithm that utilizes TI-RADS-defined ultrasound properties as ML features, (ML)TI-RADS, (4) deep learning (DL) which employed convoluted neural network algorithms to achieve prediction. For prediction of tumor (T) stage, nodal (N) stage, extrathyroidal extension, and BRAF, a tri-model that consists of radiomics, TDA, and (ML)TI-RADS was used. One hundred and four quantitative radiomics features, 476 topological algebraic-geometric features, and 5 TI-RADS features were obtained for all nodules. Linear discriminant analysis (LDA) and the Pearson correlation coefficient (0.85 threshold) were employed for feature extraction. Support vector machine (SVM) was chosen as the machine learning classifier. <h3>Results</h3> For malignancy prediction, radiomics, TDA, (ML)TI-RADS, and DL individual model achieves an accuracy of 89% (0.87 AUC), 81% (0.81 AUC), 80% (0.76 AUC), and 87% (0.92 AUC), respectively. Our quad-model achieves a 98.7% accuracy, 0.98 AUC, 96% precision, 98% sensitivity and specificity for malignancy prediction, which is significantly improved when compared to individual model (radiomics, TDA, (ML)TI-RADS: p<0.001, deep learning: p=0.002). Using the external validation dataset, the quad-model achieves a 93% accuracy (0.94 AUC) for malignancy prediction. Our tri-model yields an accuracy of 93% accuracy (0.93 AUC), 89% accuracy (0.88 AUC), 98% (0.96 AUC), and 96% (0.97 AUC) for prediction of T, N, extrathyroidal extension, and BRAF mutation, respectively. <h3>Conclusion</h3> Using routine clinical ultrasonography, we have established a robust AI platform that allows prediction of malignancy, pathological stage, and BRAF mutation for thyroid cancer with high accuracy.