Thyroid nodules, defined as localized abnormal cellular proliferations within the thyroid gland, are typically detected through palpation or imaging examinations. While most nodules are benign and often incidentally discovered during routine physical examinations or US, clinically significant nodules may cause compression symptoms that impair normal swallowing or respiratory function when exceeding certain dimensions. In thyroid tissue, the BRAF mutation is most commonly defined by the V600E variant, caused by a T→A transversion at nucleotide 1799 (exon 15). Among malignant thyroid nodules, papillary thyroid carcinoma (PTC) represents the most prevalent histological subtype. Recent advances in molecular diagnostics have identified the BRAF^V600E mutation as a specific molecular marker for PTC (1, 2). The BRAF^V600E mutation results in the constitutive activation of its encoded oncoprotein, conferring elevated kinase activity. This leads to the aberrant and persistent activation of the MAPK signaling pathway (MEK1/2 and ERK1/2) through a phosphorylation cascade. The consequent dysregulated signaling serves as the principal molecular driver promoting tumorigenesis by inducing uncontrolled cellular proliferation, inhibiting apoptosis, and facilitating abnormal differentiation and invasion of thyroid cells (3–6). The BRAF^V600E mutation in thyroid nodules is a significant indicator of malignancy and is associated with greater biological aggressiveness (7, 8), manifesting as higher postoperative recurrence rates and poorer clinical prognosis. Current clinical guidelines advocate active surveillance for low-risk nodules. Conversely, BRAF^V600E-positive cases warrant more aggressive interventions, such as surgical resection or enhanced monitoring protocols with more frequent US examinations. Precise assessment of BRAF mutation status is essential not only for definitive tumor diagnosis and effective anti-BRAF targeted therapy (e.g., vemurafenib and dabrafenib) but also for providing crucial prognostic information regarding the nodule’s biological behavior and potential aggressiveness (9, 10). While BRAF^V600E genotyping typically depends on surgical or fine-needle aspiration biopsy samples analyzed by RT-PCR or NGS, inadequate tissue or poor sample quality frequently impedes molecular diagnosis (11). Moreover, since a single biopsy or small tissue specimen only captures the genetic status of the sampled site, it fails to assess the comprehensive genetic panorama of the tumor, thereby potentially missing other critical co-mutations. Although droplet digital PCR has demonstrated high sensitivity and specificity in detecting BRAF mutations in plasma circulating cell-free DNA, its standalone diagnostic utility remains limited. In some cancer patients, the tumor burden may be insufficient to yield detectable levels of circulating tumor DNA (12). The development of imaging-based integrated models for predicting genetic mutation status, such as BRAF^V600E, represents a highly promising direction in precision medicine. The core advantage of this approach lies in its ability to overcome the major limitations of conventional molecular detection methods. It is non-invasive, thereby avoiding the risks associated with additional biopsies, including bleeding, infection, and tumor seeding. Furthermore, by capturing information from the entire tumor lesion, it has the potential to provide a more comprehensive reflection of the tumor’s global biological behavior, thereby offering a novel perspective and tool for clinical decision-making.

The Thyroid Imaging Reporting and Data System (TI-RADS) is an ultrasound-based standardized classification system designed to stratify malignancy risk in thyroid nodules and guide clinical management. Thyroid nodules classified as TI-RADS categories 1 and 2 demonstrate minimal malignancy risk, obviating the need for fine-needle aspiration or routine surveillance. A progressive increase in malignant potential was observed with ascending TI-RADS classification levels (13). CECT is an imaging technique that utilizes intravenous contrast agents in conjunction with CT scanning to evaluate tissue vascularity and structural morphology. In the diagnostic workup of thyroid nodules, high-frequency US and CECT serve as complementary imaging modalities, each with distinct advantages and limitations. US, with its superior soft-tissue resolution, is the primary modality for evaluating internal nodular architecture - such as microcalcifications, solid components, and borders - as well as vascularity. Its real-time, dynamic, and radiation-free nature makes it ideal for initial screening, risk stratification, longitudinal follow-up, and, uniquely, for guiding fine-needle aspiration biopsy. However, its diagnostic accuracy is operator-dependent, and its utility is limited in assessing retrosternal extension, deep cervical lymph node metastases, and invasion into adjacent structures. In contrast, CECT provides a comprehensive anatomical roadmap of the neck and mediastinum. It excels at delineating the spatial relationships between a nodule and critical structures like the trachea, esophagus, and major blood vessels. This capability is indispensable for evaluating extra-thyroidal extension in advanced cancer, retrosternal growth, and mediastinal lymph node metastasis, thereby providing critical information for preoperative planning. Its principal limitation, however, lies in a lower sensitivity for detecting the subtle internal features of a nodule compared to US.

Most existing studies are confined to a single imaging modality, and their predictive performance has room for improvement. The combination of US and CECT offers a synergistic approach, enabling not only precise qualitative assessment and risk stratification of nodules but also a comprehensive evaluation of their overall extent and potential invasiveness. This facilitates a co-diagnostic process that integrates anatomical localization with pathological qualification, thereby optimizing therapeutic decision-making. Therefore, this study pioneers the construction of a comprehensive predictive model by integrating multimodal features from these two routine imaging techniques. We aim to explore the correlation between BRAF^V600E mutation status and multimodal imaging features, with the expectation that such informational complementarity will yield superior predictive performance compared to any single-modality approach. The primary objective of this study is to develop and validate multimodal predictive models for preoperatively assessing BRAF^V600E mutation status. We will construct and compare three distinct models: one based on US features, another on CECT features, and an integrated model combining both. Through a comparative performance analysis, the optimal model will be selected to establish a risk stratification tool. The ultimate goal is to provide clinicians with a more reliable diagnostic framework to facilitate timely intervention for thyroid nodules with a high risk potential.

Medical imaging serves as a well-established, non-invasive diagnostic tool, offering a promising avenue for predicting genomic status. Previous studies have validated the efficacy of CT-based radiomic models in predicting KRAS, NRAS, and BRAF mutations in colorectal cancer (14–16). In the context of thyroid cancer, CT radiomics has been primarily applied to assess the risk of lymph node metastasis (17). However, compared to its successful application in the genotyping of colorectal cancer, the potential of CT radiomics for predicting key driver gene mutations in thyroid cancer, such as BRAF^V600E, remains underexplored. Concurrently, while ultrasonographic radiomics has demonstrated well-established value in differentiating benign from malignant thyroid nodules (18), its utility in predicting BRAF^V600E mutation has been limited (19, 20). Furthermore, radiomic analysis typically requires specialized software and additional computational time, preventing its seamless integration into existing clinical workflows and thus posing challenges for practical application in routine clinical settings. Therefore, the BRAF^V600E mutation prediction model developed in this study, which integrates conventional US and CECT features, offers distinct advantages over previous radiomics models. These advantages lie not only in a potential improvement in predictive accuracy but, more importantly, in its innovative multimodal approach. This integration enables complementary strengths of the two imaging modalities, thereby significantly enhancing the model’s interpretability and its acceptability to clinicians.

The core clinical value of the US combined CECT nomogram model lies in its provision of a novel decision-support tool for the refined and individualized management of thyroid nodules classified as C-TIRADS 3 and above. In terms of clinical applicability, all variables incorporated into the model are routinely and readily available from standard clinical examinations, requiring no additional costly or invasive procedures. This ensures the model’s strong generalizability and operational feasibility in diverse practice settings. The potential impact of this model on clinical decision-making is twofold. First, for nodules predicted to be at high risk of BRAF^V600E mutation, clinicians can recommend preoperative fine-needle aspiration for genetic testing or consider diagnostic surgery with greater confidence. This approach aids in mitigating the risk of underdiagnosis and delayed treatment for high-risk lesions. Second, for nodules predicted to be at low risk, the model serves as a valuable supplement to the conventional C-TIRADS classification. It provides additional evidence to support the choice of ‘active surveillance’ over immediate biopsy, thereby potentially reducing unnecessary invasive procedures along with their associated complications and healthcare costs. For clinicians, this model offers a powerful decision-making aid. It translates complex imaging features into an intuitive probability of mutation risk. When faced with nodules of diagnostic uncertainty, it empowers physicians to make more informed assessments regarding the benefits and risks of biopsy, facilitating the development of tailored management strategies. Ultimately, this tool is poised to enhance the overall precision and confidence in the clinical management of thyroid nodules.

Multiple studies have indicated a significant association between the BRAF^V600E mutation and specific ultrasonographic features (21–23). Consistent with these reports, our study identified a higher C-TIRADS category, an irregular morphology, and the presence of microcalcifications as independent risk factors for the BRAF^V600E mutation in thyroid nodules. Conversely, nodule size emerged as an independent protective factor, with smaller nodules exhibiting a higher mutational risk. These findings suggest that the BRAF^V600E mutation does not occur randomly but is closely associated with a more aggressive imaging phenotype. The underlying mechanism likely involves the malignant biological behavior driven by this mutation, which directly manifests as these characteristic sonographic changes (24). Our study demonstrated that the risk of BRAF^V600E mutation increases significantly with ascending C-TIRADS categories. This result aligns with previous research, which has also reported significant correlations between the BRAF^V600E mutation and higher risk scores in both EU-TIRADS and ACR TI-RADS systems (25, 26). Collectively, this converging evidence from different TIRADS frameworks confirms a strong link between the BRAF^V600E mutation and imaging features indicative of malignancy. It underscores the universal value of standardized sonographic categorization in the preoperative, noninvasive prediction of this pivotal molecular event. An irregular nodular morphology is a recognized sonographic marker of invasive growth behavior in thyroid cancer, which is closely linked to molecular pathways driven by the BRAF^V600E mutation. Consequently, a strong and well-defined positive correlation exists between these two features. In malignant nodules, excessively rapid tumor cell proliferation can lead to the overgrowth of fibrous tissue and blood vessels, subsequently causing calcium salt deposition and the formation of calcifications. In our study, only microcalcifications demonstrated clinical value, whereas macrocalcifications showed no statistical significance. This discrepancy may be attributed to the fact that macrocalcifications often result from dystrophic calcification within fibrotic connective tissue, which is not typically associated with malignant or invasive potential (27). The relationship between nodule size and malignancy risk remains controversial. Some studies suggest that smaller nodules (e.g., those with a diameter < 2.0 cm) may harbor a higher probability of malignancy (28, 29). For instance, Kamran et al. reported that the malignancy risk increased proportionally in nodules measuring 1.0–2.0 cm but stabilized beyond this threshold (30). In contrast, other perspectives maintain that nodule size alone is not a reliable predictor of malignancy risk (31). Our study revealed that smaller nodules were associated with a higher risk of BRAF^V600E mutation. This finding implies that the BRAF^V600E mutation may be present from the early stages of tumorigenesis, acting as an initial driver event that confers malignant potential even in small nodules (30). As a nodule enlarges, cytological heterogeneity may develop, with malignant cells coexisting with a substantial proportion of benign cells, potentially diluting the mutational signal (32). Additionally, selection bias in clinical detection could influence these observations. Smaller nodules are often selected for genetic testing or surgery due to suspicious cytological or imaging features, thereby increasing the sensitivity of mutation detection. In contrast, larger nodules are frequently resected due to compressive symptoms or cosmetic concerns, even when their cytology is benign. This difference in clinical pathways might lead to a relatively higher false-negative rate for mutation detection in larger nodules (33). In summary, no unified consensus currently exists regarding the association between nodule size and BRAF^V600E mutation or overall thyroid cancer risk. This complex relationship involves intricate tumor biological mechanisms and clinical detection biases, warranting further investigation for clarification.

An analysis of medical records from over one million patients revealed a positive correlation between the utilization of neck CT imaging and an increased incidence of thyroid cancer across different population subgroups (34). In line with the heightened diagnostic sensitivity of imaging, our study found that thyroid nodules exhibiting blurred boundaries and interrupted capsules on CECT were significantly associated with the BRAF^V600E mutation. This distinct imaging phenotype is not incidental but rather a macroscopic reflection of the specific tumor biology driven by the BRAF^V600E mutation. Mutated tumor cells secrete a variety of proteases that degrade the extracellular matrix, facilitating invasive growth into the surrounding thyroid parenchyma. Concurrently, this process induces the formation of aberrantly abundant neovessels at the tumor periphery (35). On CECT, these pathological changes manifest as an enhancing, invasive rim, collectively contributing to the ‘blurred boundaries’. When these highly aggressive mutant tumors expand, they directly invade and breach the thyroid capsule. This pattern stands in sharp contrast to the expansive growth typically observed in benign nodules. Even when significantly enlarged, benign nodules are often surrounded by a fibrous capsule, maintaining a clear demarcation from adjacent tissues (36).

The multimodal model developed from the aforementioned imaging features holds significant value for predicting BRAF^V600E mutation risk in thyroid nodules categorized as C-TIRADS 3 and above. Compared to models based solely on the US or CECT, the integrated model demonstrated superior performance across all evaluation metrics: its calibration curve was closest to the ideal line, it achieved the highest AUC, it yielded a higher net benefit across a wider range of threshold probabilities in decision curve analysis, and it had the lowest Brier score. These results collectively indicate optimal performance in calibration, discrimination, clinical utility, and overall accuracy, underscoring the considerable promise of fusing US and CECT features for preoperative, non-invasive genotyping.

A key consideration in this study was the inclusion of TI-RADS category 3 nodules. We recognize that current guidelines classify these nodules as having a low malignancy risk and typically recommend active surveillance rather than immediate intervention. The primary objective of our prediction model is not to overturn this standard management strategy but to address a more nuanced clinical scenario. Specifically, it is intended to aid in decision-making when the management of certain TI-RADS 3 nodules—for instance, those that are large (e.g., >2 cm), cause patient anxiety, or exhibit growth during follow-up—is being reconsidered for potential FNA. At this critical juncture, our model integrates ultrasonographic and CECT features to provide a probability assessment for the presence of the high-risk BRAF^V600E mutation. A high-risk prediction from the model could offer evidence-based support for BRAF mutation testing in such nodules, potentially leading to earlier definitive diagnosis and intervention, thereby avoiding possible diagnostic delays. Conversely, a low-risk prediction could reinforce clinician and patient confidence in continuing with active surveillance. Future studies should focus on validating the clinical utility and cost-effectiveness of this model within this specific TI-RADS category 3 subgroup for guiding precision FNA and genetic testing strategies.

This study has several limitations that should be acknowledged. First, the sample size was relatively limited, with a moderate class imbalance between the BRAF^V600E mutant (n = 61) and wild-type (n = 263) groups. This imbalance may introduce selection bias and potentially affect the model’s predictive performance. Despite initial efforts to mitigate class imbalance via oversampling and undersampling, both methods resulted in models with compromised performance relative to the model developed from the unaltered data (Supplementary Materials 1-3). Given these findings, data resampling was therefore abandoned in the final analytical approach. To mitigate the impact of class imbalance, we employed several strategies: we prioritized the retention of features strongly associated with the mutation, applied L1 regularization for feature selection to eliminate redundant variables, and implemented a weighted logistic regression approach that assigned a higher misclassification cost to samples in the mutant group. This encouraged the model to focus more on learning from the minority class during fitting. These measures were taken to objectively reflect the model’s performance across both nodule types. Second, this study has a single-center, retrospective design. All enrolled cases were recruited from a single institution, and patients who declined FNA or BRAF^V600E genetic testing were excluded, which may introduce selection bias. Furthermore, the model’s performance may be influenced by our center-specific patient demographics, imaging protocols, and interpretation conventions, potentially limiting the generalizability of the developed nomogram. To address these limitations, we are currently expanding the sample size at our center and have initiated a collaborative multi-center study with other peer-level teaching hospitals in the region. We aim to establish a larger, more balanced cohort to facilitate external validation, thereby further verifying the reliability and generalizability of our findings. Finally, the FNA procedure is subject to sampling variability. The accuracy of BRAF^V600E genotyping may be compromised by false-negative results, which can arise from intratumoral heterogeneity or inadequate sampling of the mutational component.

In conclusion, the multimodal imaging model based on US and CECT has better predictive performance in predicting BRAF^V600E gene mutations of thyroid nodules of TI-RADS category 3 and above compared with the single imaging feature model. This validated multimodal model could potentially reduce costs by improving risk stratification and reducing clinically unnecessary genetic testing. In addition, clinical decision support can be enhanced by individualized estimates of the probability of malignancy that can inform diagnosis and treatment pathways. It is beneficial to optimize the allocation of resources in the medical system while maintaining the accuracy of diagnosis.