背景和目的

甲状腺结节是最常见的内分泌疾病，其中恶性结节约占10%。准确诊断恶性结节是制定治疗方案的首要步骤。然而，目前鉴别甲状腺良恶性结节的超声分层系统仍存在诸多的缺陷。因此，应用更多超声模态特征提高恶性结节的诊断率显得尤为迫切。

分化型甲状腺癌（differentiated thyroid carcinoma, DTC）占甲状腺癌的80%以上，颈部淋巴结（lymph nodes, LNs）转移是影响DTC患者预后的重要因素之一。目前，超声仅能筛查出不足30%的转移LNs。超声引导下LNs穿刺活检对操作者技术要求高，位置较深的LNs常难以获得有效的穿刺标本。因此，迫切需要一种新方法提高DTC转移LNs的术前检出率，这对手术范围的确定和患者预后均具有重要影响。

DTC中最主要的病理类型为甲状腺乳头状癌（papillary thyroid carcinoma, PTC），约60%以上的PTC合并BRAF^V600E基因突变。BRAF^V600E突变的PTC患者出现复发和转移的风险高。穿刺进行基因检测是术前诊断BRAF^V600E突变的主要方法，但是其为有创检查且价格昂贵。同时，对于10mm以下的PTC（即甲状腺微小乳头状癌，PTMC）是否应进行穿刺尚有争议，这使术前识别高侵袭性PTMC变得更加困难。因此，从超声图像中筛选与BRAF^V600E突变相关的特征、建立无创的影像预测模型将有望解决目前的困境。

人工智能（artificial intelligence, AI）可通过计算机提取感兴趣区域的图像信息，将主观的定性诊断转变为客观的定量分析。影像组学作为AI分析方法之一，主要通过机器学习算法从超声图像中高通量提取人眼无法识别的定量特征。超声影像组学应用于甲状腺良恶性结节鉴别、PTC转移LNs诊断中的可行性已被既往研究初步证实，但是融合多模态超声成像，包括灰阶超声、彩色多普勒成像（color doppler flow imaging, CDFI）和超声造影（contrast-enhanced ultrasound, CEUS）等影像组学分析的诊断价值尚未得到充分验证。另外，深度学习算法在甲状腺疾病诊断中的应用也得到了较大的发展，由于其更强的图像捕捉和信息提取能力，深度学习模型在处理动态超声视频（如CEUS）中表现出巨大潜能。

基于以上认识，针对DTC诊断、BRAF^V600E突变预测和颈部转移LNs识别的三大临床问题，本研究拟探索应用不同AI方法对灰阶超声、CDFI和CEUS图像特征的提取效果，研究超声影像组学模型在DTC各类诊断问题中的应用价值。本研究的目标是：（1）探索PTMC灰阶单模态超声影像组学参数与BRAF^V600E突变之间的关系，建立基于超声影像组学特征的列线图模型并评估其诊断效能；（2）对DTC颈部LNs灰阶超声和CDFI图像进行双模态影像组学信息提取和筛选，联合影像组学分数、临床信息和超声征象建立诊断转移LNs的综合模型，并评估其应用价值；（3）构建基于视频视觉Transformer（ViViT）的甲状腺CEUS诊断模型，并应用于甲状腺结节良恶性鉴别和PTC BRAF^V600E突变预测中，并验证其可行性和有效性。

方法

（1）回顾性纳入符合诊断标准的PTMC，获取结节最大切面的灰阶图像并记录结节的超声特征，应用MaZda软件提取影像组学特征。应用LASSO（least absolute shrinkage and selection operator）回归分析筛选出与BRAF^V600E突变相关的影像组学特征并计算影像组学分数，再采用Logistics回归分析确定与BRAF^V600E突变相关的独立预测因素，建立基于超声影像组学的列线图模型，并将其与传统超声特征模型对比。

（2）回顾性纳入两个医学中心确诊为DTC且可疑颈部LNs肿大的患者，记录临床信息和LNs超声特征。根据欧洲甲状腺协会（European Thyroid Association, ETA）分级将LNs分为可疑恶性、正常和不定性LNs。获取LNs最大切面的灰阶和CDFI图像，应用Pyradiomics软件提取双模态影像组学特征，采用LASSO和Logistics回归分析确定转移LNs的独立预测因素。构建基于超声影像组学的LNs诊断模型且应用于ETA分类为不定性LNs的诊断中，并在多中心数据中验证其诊断效能。

（3）回顾性纳入良恶性诊断明确且有BRAF^V600E基因诊断结果的甲状腺结节，收集CEUS动态视频，同时记录超声和CEUS特征。将CEUS数据以3:1:1的比例分为训练集、验证集和测试集。构建并验证基于ViViT的CEUS诊断模型，再将测试集数据代入模型进行分类诊断。将ViViT-CEUS甲状腺结节良恶性诊断模型的效能分别与二维卷积神经网络（two dimensions-convolutional neutral networks, 2D-CNN）、三维卷积神经网络（three dimensions-convolutional neutral networks, 3D-CNN）和不同年资的超声医师进行对比。再将ViViT-CEUS预测 BRAF^V600E突变的效能与2D-CNN、3D-CNN和传统影像组学模型进行对比，综合评估ViViT-CEUS模型的诊断价值。

结果

（1）共纳入328例PTMC，分为训练组（n=232）和验证组（n=96）。从灰阶单模态超声图像中筛选出4个影像组学特征并计算影像组学分数。Logistics回归显示影像组学分数、结节成分和纵横比≥1是PTMC BRAF^V600E突变的独立预测因素。基于上述三个参数建立的诊断模型具有良好的校准能力，且超声影像组学诊断模型比传统超声特征诊断模型具有更高的预测能力（训练集AUC 0.685 vs 0.592，验证集AUC 0.651 vs 0.622）和更高的临床净获益。

（2） 医学中心1共纳入611例淋巴结，分为训练集（n=428）和内部测试集（n=183），医学中心2共115例作为外部验证集。从灰阶和CDFI双模态超声图像中提取并筛选出37个影像组学特征，计算影像组学分数。Logistics回归显示年龄、DTC位置、LNs分区、LNs短径、淋巴结门消失、囊性变、局部高回声、血流信号和影像组学分数是转移LNs的独立预测因素。基于上述因素建立的诊断模型显示出良好的校准能力。加入超声影像组学的诊断模型比传统超声诊断模型具有更高的诊断能力（训练集AUC 0.871 vs 0.848，内部测试集AUC 0.804 vs 0.803，外部验证集AUC 0.939 vs 0.921）。且新型模型可显著提高ETA不定性LNs中转移LNs的检出率（AUC 0.868 vs 0.793）。

（3）共纳入428例CEUS视频完整的甲状腺结节，其中良性69例、恶性359例。恶性结节均为PTC , 其中BRAF^V600E突变267例，未突变92例。ViViT-CEUS模型鉴别甲状腺结节良恶性的敏感性、特异性、阳性预测值、阴性预测值和准确性分别为93.87%、85.51%、97.12%、72.84%和92.52%。AUC为0.897，显著高于2D-CNN、3D-CNN模型以及三位超声医师（DeLong 检验p值均小于0.001）。ViViT-CEUS模型预测PTC BRAF^V600E的敏感性、特异性、阳性预测值、阴性预测值和准确性分别76.78%、77.17%、90.71%、53.38%和76.88%。AUC为0.77，显著高于2D-CNN、3D-CNN和传统影像组学模型（DeLong 检验p值均小于0.05）。

结论

（1）灰阶单模态超声影像组学特征与PTMC BRAF^V600E突变显著相关，联合超声影像组学分数和常规超声特征的模型具有预测PTMC BRAF^V600E突变的潜力。

 （2）从灰阶超声和CDFI双模态超声图像中可筛选出与DTC转移LNs显著相关的影像组学特征。联合双模态超声影像组学分数、临床信息和超声特征构建的转移淋巴结的新型诊断模型具有较高的诊断准确性，且在ETA不定性LNs中，新型影像组学模型可显著提高转移LNs的检出率。

（3）基于ViViT构建的深度学习模型从时域及空域对CEUS进行了有效信息提取。ViViT-CEUS模型实现了对甲状腺良恶性结节的准确分类，与2D-CNN、3D-CNN和超声医师相比均具有更高的诊断效能。ViViT-CEUS模型也提高PTC BRAF^V600E突变的诊断准确性，为利用动态超声图像识别PTC分子特征提供了实验基础。

Background

Thyroid nodules are the most common thyroid disease with about 10% malignant cases. The current ultrasound (US)-based stratification systems for diagnosing thyroid malignancy still have limitations. Therefore, adding multiple US modalities to improve the diagnostic rate of malignant nodules is necessary.

Differentiated thyroid carcinoma (DTC) accounts for more than 80% of thyroid cancers. Approximately 20-50% of DTC patients have lymph node (LN) metastasis. US examination can only detect less than 30% of metastatic LNs. Fine-needle aspiration (FNA) for LNs requires high technical skills, and sometimes it’s hard to obtain effective samples. Therefore, there is an urgent need for a new method to improve the detection rate of DTC metastatic LNs, as it has a significant impact on the scope of surgery and prognosis.

The most common pathological type of DTC is papillary thyroid carcinoma (PTC). More than 60% of PTC have BRAF^V600E gene mutation, which indicates greater risk of recurrence and metastasis. Preoperative genetic testing through FNA is the primary method to diagnose BRAF^V600E mutation. However, it is an invasive procedure with high cost. Additionally, there is controversy about whether FNA should be performed for PTCs less than 10mm (papillary thyroid microcarcinoma, PTMC), making it more challenging to identify highly invasive PTMC before surgery. Exploring US features associated with the BRAF^V600E mutation and establishing a non-invasive prediction model holds promise for addressing the current dilemma.

Artificial intelligence (AI) utilizes computer technology to extract information from medical images，transforming subjective qualitative diagnoses into objective quantitative analysis. Radiomics, as one of the AI methods, primarily use machine learning algorithms to extract quantitative features from US images. The feasibility of US-based radiomics in thyroid diagnosis has been preliminarily confirmed in previous studies. However, the diagnostic value of radiomics based on multimodal US imaging, including grayscale US, color doppler flow imaging (CDFI), and contrast-enhanced ultrasound (CEUS), has not been extensively validated. Additionally, the application of deep learning algorithms in thyroid disease diagnosis has witnessed significant advancements. Deep learning models, with their stronger capabilities of capturing and extracting information, demonstrate tremendous potential in handling dynamic US videos (e.g., CEUS).

Based on the above research background, this study aims to address three major issues including DTC diagnosis, BRAF^V600E prediction and metastatic LN recognition. Specifically, we intend to use different AI methods to extract features from grayscale US, CDFI and CEUS, respectively. The specific objectives as follows: (1) To explore the relationship between BRAF^V600E mutation and radiomic parameters from grayscale US images, establish a nomogram based on radiomic features, and evaluate its diagnostic performance; (2) To extract radiomic information from both grayscale and CDFI images of DTC LNs, develop a model combining Rad-score, clinical information, and US features for diagnosing metastatic LNs (3) To construct a thyroid CEUS diagnostic model based on video vision transformer (ViViT) algorithm, applying it to distinguish malignant thyroid nodules and predict BRAF^V600E mutation of PTC.

Methods

(1) We retrospectively enrolled PTMC cases and obtained grayscale US images of the largest cross-section of the nodules. All the US features were recorded. MaZda software was used to extract radiomic features. Then we applied LASSO (least absolute shrinkage and selection operator) regression to select radiomic features associated with the BRAF^V600E mutation and calculated Rad-score. Logistic regression analysis was used to determine independent predictive factors of BRAF^V600E mutation. Nomogram based on Rad-score was established and its diagnostic ability was compared with that of conventional US models.

(2) DTC cases with cervical lymphadenopathy in two medical centers were retrospectively enrolled. LNs were classified into suspicious malignant, normal, and indeterminate categories according to the European Thyroid Association (ETA) classification. Grayscale and CDFI images of the LNs were obtained, and Pyradiomics software was used to extract bimodal radiomic features. LASSO and logistic regression were applied to determine independent predictive factors for metastatic LNs. A nomogram based on bimodal Rad-score was developed and applied to recognize metastatic LNs in ETA indeterminate LNs.

(3) We retrospectively collected benign and malignant thyroid nodules with CEUS videos. All the malignant nodules had BRAF^V600E diagnostic results. The CEUS data were divided into training, validation and testing sets in a ratio of 3:1:1. A CEUS diagnostic model based on video vision transformers (ViViT) was constructed and validated. Then we used testing data for diagnosis. The diagnostic performance of the ViViT-CEUS model to distinguish malignant thyroid nodules was compared with that of two-dimensional convolutional neural networks (2D-CNN), three-dimensional convolutional neural networks (3D-CNN), and ultrasound physicians. The ability of ViViT-CEUS model to predict BRAF^V600E mutation was also be compared with 2D-CNN, 3D-CNN, and radiomics models.

Results

(1) 328 PTMCs were enrolled and divided into training (n=232) and test (n=96) cohorts. Four radiomic features were selected from grayscale US images. Rad-score, composition and aspect ratio were independent predictive factors of BRAF^V600E status. The nomogram incorporating three variables showed good calibration. Compared with conventional US model, the radiomics nomogram showed better diagnostic ability (training set AUC 0.685 vs 0.592, testing set AUC 0.651 vs 0.622), and superior clinical applicability.

(2) 611 LNs from center 1 were randomly divided into training (n=428) and internal testing (n=183) groups. 115 cases from center 2 were external testing group. From grayscale US and CDFI, LASSO regression screened 37 radiomics features and calculated the Rad-score. Age, DTC location, LN region, LN short axis, hilum absence, cystic degeneration, focal hyperecho, blood flow and Rad-score were independent predictors of LN metastasis. The diagnostic model based on the above factors showed great calibration ability. The radiomics model had better diagnostic performance than the traditional US model (training AUC 0.871 vs 0.848, internal testing AUC 0.804 vs 0.803, external testing 0.939 vs 0.921). Furthermore, the novel radiomics model could significantly improve the diagnostic accuracy of metastatic LNs in ETA indeterminate LNs (AUC 0.868 vs 0.793).

(3) 428 thyroid nodules were included, consisting of 69 benign and 359 malignant nodules. All malignant nodules were PTC with 267 BRAF^V600E mutations and 92 non-mutated cases. The ViViT-CEUS model for diagnosing benign and malignant thyroid nodules achieved SE, SP, PPV, NPV and ACC of 93.87%, 85.51%, 97.12%, 72.84%, and 92.52%, respectively. The AUC of ViViT-CEUS was 0.897 which significantly higher than that of 2D-CNN, 3D-CNN models and all the physicians (DeLong test, all p-values<0.001). The ViViT-CEUS model for predicting BRAF^V600E achieved SE, SP, PPV, NPV and ACC of 76.78%, 77.17%, 90.71%, 53.38%, and 76.88%, respectively. The AUC the model was 0.77, which greatly higher than that of 2D-CNN, 3D-CNN and radiomics models (DeLong test, all p-values < 0.05).

Conclusions

(1) Grayscale US-based radiomic features are significantly associated with PTMC BRAF^V600E mutation. The model combining radiomic features and conventional US signs shows potential for predicting BRAF^V600E in PTMC patients. (2) The bimodal radiomics features from grayscale US and CDFI images shows significant correlation with DTC metastatic LNs. The diagnostic model combining bimodal US Rad-score, clinical information and ultrasound signs displays favorable capability, and it can greatly improve the detection rate of metastatic LNs in ETA indeterminate LNs. (3) ViViT algorithm can extract effective information from CEUS videos. This ViViT-CEUS model achieves high accuracy to identify malignant thyroid nodules, surpassing the diagnostic ability of 2D-CNN, 3D-CNN and US physicians. It also improves the diagnostic accuracy of PTC BRAF^V600E mutations, providing an experimental basis for identifying molecular features based on dynamic CEUS videos.