深度学习
医学
超声波
突变
恶性肿瘤
人工智能
放射科
计算机科学
甲状腺
生物信息学
文本挖掘
病理
超声科
甲状腺癌
计算生物学
甲状腺癌
作者
Jiajia Tang,Enguang Sui,Yan Tian,Ruyu Liu,Xinyi Liu,Jianlin Wang,Bo Zhang
标识
DOI:10.1038/s41598-026-77788-9
摘要
To develop a video vision transformer (ViViT) model based on thyroid contrast-enhanced ultrasound (CEUS) video, applying it to distinguish malignant thyroid nodules and predict BRAFV600E mutation in papillary thyroid carcinoma (PTC). We retrospectively collected thyroid nodules with CEUS videos and divided them into training, validation, and testing sets in a ratio of 4:3:3. The ViViT-CEUS diagnostic model was developed and validated. Its performance was compared with two-dimensional convolutional neural networks (2D-CNN), three-dimensional convolutional neural networks (3D-CNN) and traditional radiomics models. Radiologists with different levels of clinical experience were served as the comparator group. 428 thyroid nodules were included (69 benign, 359 malignant). All malignant nodules were PTC, comprising 267 BRAFV600E mutated and 92 non-mutated cases. The AUC of ViViT-CEUS for diagnosing malignancy was 0.938, which was significantly higher than that of 2D-CNN (AUC = 0.767), 3D-CNN (AUC = 0.790), and radiologists (AUC range : 0.617–0.815). The AUC of ViViT-CEUS for predicting BRAFV600E was 0.865, outperforming 2D-CNN (AUC = 0.704), 3D-CNN (AUC = 0.597), and traditional radiomics models (AUC = 0.682). The ViViT-CEUS model demonstrated high accuracy in identifying malignant thyroid nodules and improved prediction of PTC BRAFV600E mutations compared to alternative methods, providing a basis for identifying molecular features based on dynamic CEUS videos.
科研通智能强力驱动
Strongly Powered by AbleSci AI