AI‐BRAFV600E: A deep convolutional neural network for BRAFV600E mutation status prediction of thyroid nodules using ultrasound images

Background: The BRAFV600E mutation is a valuable indicator for thyroid cancer diagnosis. This study aimed to develop a deep convolutional neural network (DCNN) model based on ultrasound images to predict the BRAFV600E mutation status of thyroid nodules. Methods: The ultrasound images were obtained from four hospitals between January 2017 and January 2022. We trained and validated the DCNN model based on the primary set from center 1 (979 images, 528 patients). The DCNN network consists of Conv block, Downsample block, Gaussian error linear unit, Global Average Polling, and Full Connected. The predictive performance of this model was evaluated by using areas under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity in four independent test sets from center 1 to center 4 (531 images, 282 patients). Heatmaps were used to visualize the most predictive regions of each image. Specimens obtained through fine‐needle aspiration or surgery were used to detect the BRAFV600E mutation. Results: The DCNN model achieved encouraging predictive performance by fivefold cross‐validation (AUC 0.95) in the primary set. This performance was further confirmed in the independent internal test set (AUC 0.93) and three independent external test sets (AUC 0.84–0.88). The deep learning score revealed significant differences between BRAFV600E‐mutant and BRAFV600E‐wild‐type groups (all test sets p < .001). The heatmaps visualized the most predictive region located inside or alongside the thyroid nodules. Conclusion: A DCNN model with encouraging predictive performance was developed based on ultrasound images to predict the BRAFV600E mutation status of thyroid nodules.