Multi-Task Network Guided by Ultrasound Features for Predicting BRAFV600E Mutation Status in Thyroid Ultrasound Images

Thyroid cancer is recognized as one of the most prevalent malignancies worldwide, with its incidence often linked to the BRAF^V600E mutation, a mutation in the BRaf protooncogene serine/threonine kinase (BRAF). The conventional method for detecting this mutation involves invasive fine-needle aspiration, highlighting the urgent need for a noninvasive alternative. This study aims to establish a predictive framework for BRAF^V600E mutation status in thyroid cancer by leveraging the correlation between BRAF^V600E and various ultrasound image features. The goal is to introduce a noninvasive technique for determining the mutation status, thus advancing thyroid cancer diagnostics. The investigation thoroughly examined ultrasound images of 3310 thyroid nodules, including 2115 instances of the BRAF^V600E mutation, using a dataset approved by the Ethics Committee of Tianjin Medical University Cancer Institute and Hospital. A deep learning-based multitask model was developed and trained on a collection of 2718 images, which were marked by imbalanced feature labels. The model was then rigorously tested on a balanced set of 592 images to determine the mutation status. Using advanced deep learning techniques, the study designed a multitask learning model proficient in predicting the presence of the BRAF^V600E mutation. This model utilized ultrasound characteristics such as composition, echogenicity, margin, echogenic foci, and shape. The model combines methods for local and global feature extraction, selection, and fusion. It begins by deriving feature representations from the ultrasound characteristics of thyroid nodules via multitask learning and then merges these features to pinpoint the signature representation indicative of the BRAF^V600E mutation. The code is publicly available at https://github.com/xuyansheng07/MTL_BRAFV600E. The model exhibited significant predictive performance, achieving an accuracy rate of 92.91%, a sensitivity of 97.94%, and a specificity of 83.25%. Additionally, relationship exploration experiments conducted in this study meticulously explored the connection between gene mutations and ultrasound features, highlighting the critical role of echogenic foci features in predicting the BRAF^V600E status. This study proposes a noninvasive method for predicting the BRAF^V600E mutation status in thyroid nodules. The findings not only demonstrate the high predictive accuracy of the model but also highlight the importance of echogenic foci in determining mutation status. The introduction of this noninvasive predictive framework opens new avenues for future research.