<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 papillary thyroid carcinoma (PTC) and is associated with poor prognosis. The purpose of this study was to establish a multimodal artificial intelligence (AI) ultrasound platform that consists of radiomics, topological data analysis (TDA), TI-RADS features, and deep learning (DL) to predict malignancy and pathological outcome in patients with papillary thyroid cancer. <h3>Materials/Methods</h3> Between 2010 and 2021, 464 patients (511 nodules) underwent fine needle biopsy at our institution. The median age was 49 years. The dataset was divided into an internal training (103 malignant, 259 benign nodules), an internal validation (51 malignant and 98 benign nodules), and an external validation (270 malignant and 50 benign nodules) datasets. High-resolution ultrasonography (sagittal and transverse views) was performed preoperatively for all patients. We employed multimodal machine learning (ML) methods: radiomics (a high-throughput quantitative analysis of image features), TDA (a rapidly growing ML pipeline that analyzes geometric relationships between data points in images), and deep learning (where algorithms run data through layers of neural network to achieve prediction). One hundred and four quantitative radiomics features, 476 topological algebraic-geometric features), and TI-RADS features were obtained for all patients. Linear discriminant analysis (LDA) along with 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, TI-RADS, and deep learning model achieves an accuracy of 88.7% (0.87 AUC), 81.5% (0.81 AUC), 80% (0.76 AUC), and 87.4% (0.92 AUC), respectively. Our multimodal AI platform which utilizes radiomics, TDA, TI-RADS, and deep learning achieves a 98.7% accuracy (0.99 AUC), which is significantly improved when compared to individual model (radiomics, TDA, TI-RADS: p<0.001, deep learning: p=0.002). On the external validation dataset, our model achieves a 91.4% accuracy (0.94 AUC) for malignancy prediction. Using multimodal that consists of radiomics, TDA, and TI-RADS, an accuracy of 93% accuracy (0.93 AUC), 89% accuracy (0.88 AUC), and 98% (0.96 AUC) is achieved for T, N, and extrathyroidal extension, respectively. Employing a 5-fold cross validation, we achieve an accuracy of 96% (0.97 AUC) for predicting BRAF mutation. <h3>Conclusion</h3> We established a robust AI platform for predicting thyroid malignancy, pathological stage, and BRAF mutation using routine ultrasonography. This low-cost reproducible approach would allow developing of personalized treatment planning, especially in countries where resources are limited.
Read full abstract