Papillary Thyroid Carcinoma (PTC) is the most common type of thyroid cancer, constituting 85-90% of malignant thyroid tumors. Although PTC progresses slowly with a generally favorable prognosis, the onset of cervical lymph node metastasis in patients can significantly increase the risk of recurrence and distant metastasis, ultimately leading to potential mortality. 2015 American Thyroid Association Management Guidelines for Adult Patients with Thyroid Nodules and Differentiated Thyroid Cancer emphasized that the number of cervical lymph node metastases is a crucial factor in assessing the recurrence risk of thyroid cancer (1). An increase in the number of lymph node metastases corresponds to a poorer clinical outcome for the patient, with a consequent reduction in the 5-year survival rate (2–4).

In the rapidly advancing realm of medical imaging, two principal ultrasonography techniques have emerged as key in predicting cervical lymph node metastasis in thyroid cancer. The first method meticulously examines the primary tumor, while the second method assesses suspicious lymph nodes. The assessment of suspicious lymph nodes to gauge the aggressiveness of thyroid cancer is a direct strategy, but the intricate anatomy of the thyroid gland, coupled with imaging technology limitations, turns preoperative ultrasonography of cervical lymph nodes into a complex task that frequently obstructs the swift identification of suspicious nodes. Therefore, the bulk of research, informed by a pragmatic approach, is derived from studies of the primary tumor, investigating attributes closely linked to the spread of cancer to cervical lymph nodes (5, 6).

The recent technological renaissance has fostered the ascent of radiomics, transcending traditional paradigms of medical imaging analysis. Algorithms meticulously mine imaging data, unveiling hidden information and enabling a comprehensive evaluation of tumor heterogeneity. This forms a foundational framework for the development of precise diagnostic and treatment models, reinforcing the pillars of clinical decision-making. Deep learning stands at the forefront of this innovation, threading significant breakthroughs in computer vision into the fabric of Artificial Intelligence (AI). This technological wonder is extensively applied in medical imaging for tasks such as segmentation, localization, detection, and image fusion, thus elevating the diagnostic precision for pathological changes. Deep learning differs from traditional machine learning—which requires intensive image preprocessing and manual feature identification—by skillfully utilizing raw pixel values from images as input and iteratively refining its models through training (7). However, to our knowledge, there has yet to be any research employing multi-modal ultrasonic radiomics data to develop corresponding deep learning models for evaluating the lymph node status of primary lesions (8, 9).

This study presents DualSwinThyroid, a deep learning classification model meticulously designed for evaluating the invasiveness of thyroid nodules. The model’s ‘Dual’ structure processes multi-modal data, and its ‘Swin’ element utilizes the Swin-Transformer’s (10) advanced image processing capabilities for thorough analysis and extraction of features from ultrasound images. ‘Thyroid’ in its name highlights the model’s specific application to thyroid nodule assessment. DualSwinThyroid is predicated on the Swin-Transformer’s solid framework and is finely calibrated to harness not just its imaging strengths but also to Integrate clinical and ultrasonic data characteristics. Such integration sharpens diagnostic accuracy and enhances efficiency, leading to more targeted and evidence-based treatment plans for patients. Additionally, DualSwinThyroid transcends the conventional binary classification of nodules, introducing a tripartite categorization that corresponds with specific therapeutic approaches and provides a clear framework for complex clinical decision-making processes, aiding in the navigation of diverse treatment alternatives.

In Papillary Thyroid Carcinoma (PTC), cervical lymph node metastasis is a common occurrence, with an estimated 40-90% of PTC patients potentially experiencing such metastasis (12, 13). In 2015, the American Thyroid Association issued management guidelines for adult thyroid nodules and differentiated thyroid cancer patients, highlighting the degree of cervical lymph node metastasis as a significant indicator of thyroid cancer recurrence. To curb the rapid progression of PTC, a crucial step is early identification of tumors with metastatic potential in clinically diagnosed PTC patients. Surgical removal is regarded as the primary treatment modality (14). For patients at risk of suspected lymph node metastasis, prophylactic central or lateral neck dissection is recommended. When performing prophylactic lymph node dissection and total thyroidectomy, the increased risk of postoperative complications, especially hypoparathyroidism, must be considered. The study by Henry et al. (15) reported that central neck lymph node dissection could escalate the risk of permanent hypoparathyroidism from 0% to 4%. However, research by Nixon I J et al. indicates that reoperation post PTC recurrence is relatively challenging, significantly elevating surgical complications and impacting the quality of life of patients (16). Hence, early identification of cervical lymph node metastasis in PTC not only aids in clinically selecting the appropriate surgical plan and scope, reducing the occurrence of postoperative complications, but also in minimizing recurrence risks, averting secondary surgeries, and proactively improving prognosis. Nonetheless, the sensitivity of solely relying on ultrasonographic characteristics to indicate lymph node metastasis remains insufficient, with some lymph node metastases exhibiting unremarkable ultrasonic features - a common scenario in clinical practice. Studies also revealed a mere 33% sensitivity of ultrasound in detecting central lymph node metastasis (17). Preoperative neck ultrasonography is inevitably influenced by inter-observer variability, thereby rendering the diagnostic outcome lacking in certain objectivity.

This study conducted a thorough investigation into the independent risk factors for cervical lymph node metastasis in Papillary Thyroid Carcinoma (PTC), analyzing 3,652 multi-modal ultrasonographic images and data from 299 patients. The results underscore the need for increased attention to cervical lymph node metastasis and recurrence risk, particularly in male patients, those aged 45 or younger, with thyroid cancer nodules larger than 1cm, nodules at the lower pole and center of the gland, presence of calcification, and suspicious lymph nodes detected by ultrasonography. To improve the accuracy of predicting cervical lymph node metastasis in PTC, the study introduced the DualSwinThyroid Model, a deep learning tool combining ultrasonographic images with clinical data. The model achieved a high AUC of 0.924, with an accuracy rate of 96.3%, and commendable sensitivity and specificity, highlighting its potential as an effective non-invasive assessment tool for lymph node involvement in PTC. Model performance is detailed in Table 3 .

Earlier research has typically hinged on machine learning and statistical methods, analyzing image, clinical, and ultrasonic features in isolation, without an in-depth approach (18–23). Luchen Chang et al. employed deep learning, alongside ultrasonic and clinical data, to develop a composite nomogram for predicting central lymph node metastasis in PTC patients, using grayscale images for radiomics and six related features. However, the absence of multi-modal ultrasonic data and additional correlating factors slightly impeded the predictive accuracy (22). Fu Li et al. achieved promising results in forecasting cervical lymph node metastasis using conventional machine learning models (19). These models, however, sometimes fail to discern complex patterns in large datasets, particularly with fluctuating data distributions or noise, which can hinder their generalization. Additionally, they depend on manual feature processing, which can compromise performance if not done meticulously. Notably, most current studies focus on merely detecting cervical lymph node metastasis, which affects the precision of choosing surgical methods and predicting patient outcomes. Details on comparative studies can be found in Table 4 .

The DualSwinThyroid model introduces a refined thyroid nodule management method by categorizing them into three different levels to enhance clinical decision-making accuracy. For Class I nodules without lymph node metastasis, unilateral lobectomy and isthmectomy are advised as the primary surgical path. In Class II cases with less than five metastatic lymph nodes, the model recommends total or near-total thyroidectomy and ‘selective’ cervical lymph node dissection, in conjunction with clinical judgment. The term ‘selective’ cervical lymph node dissection, as used here, refers to a process where the surgeon, integrating systematic preoperative evaluation with intraoperative biopsy pathology, determines the necessity of lymph node dissection and identifies the specific regions for dissection, thereby minimizing the risk of secondary surgical interventions. For Class III nodules with more than five metastatic lymph nodes, total or near-total thyroidectomy and extensive cervical lymph node dissection are advised, which is more extensive than standard elective central neck dissection, especially in cases of widespread metastasis. Physicians expand the scope of dissection based on a comprehensive assessment of other clinical indicators. This recommendation balances comprehensive treatment and the risk of overtreatment. The DualSwinThyroid model’s multi-tiered classification not only demonstrates diagnostic accuracy but also conforms to contemporary medical guidelines, enhancing patient-centered surgical decision-making.

Throughout the development of our model, several key findings emerged. Initially, it was apparent that the accuracy of predictions using single-modality data alone was relatively low, with image data outperforming clinical and ultrasonic data, highlighting the importance of imaging in predictive models. In the training dataset, accuracy confidence was higher for transverse and longitudinal images than for color Doppler and elastography. Yet, this pattern shifted in the testing phase, where color Doppler and elastography images achieved greater accuracy, revealing variable performance across imaging types during different phases. Moreover, integrating four types of images—transverse, longitudinal, color Doppler, and elastography—with clinical and ultrasonic data provided the most accurate predictions for individual cases. Adding more images did not improve but rather slightly reduced the predictive performance. These findings have been instrumental in fine-tuning the model and determining the most effective imaging techniques to increase accuracy, setting a course for future advancements in model enhancement.

This study successfully developed a multi-modal ultrasound radiomics deep learning model for predicting cervical lymph node metastasis in PTC. It aims to leverage machine learning to identify features related to metastasis, thereby improving surgical decision-making in PTC. However, the work is retrospective and exploratory in nature. Moreover, its scope is limited by its single-institution design, which might not fully represent a wider patient population. Currently, the final pathological assessment of the extent of lymph node metastasis still relies on the thoroughness and accuracy of surgical removal. While the prediction model offers a potential foundation for clinical decision support, its benefits are yet to be confirmed in prospective clinical trials. Therefore, the anticipated advantages, such as reducing lymph node dissections, financial burden, and supporting emerging practitioners, are promising but need further validation in various clinical settings. Future efforts will focus on expanding data sources and rigorously validating the model’s clinical utility in different institutions.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by First Hospital of Shanxi Medical University. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because retrospective nature of the study: if the study is retrospective, analyzing previously collected data, written informed consent may not be required as the data is often de-identified and does not pose a risk to the privacy of the individuals involved.

QL: Funding acquisition, Investigation, Methodology, Writing – original draft, Writing – review & editing. YL: Data curation, Methodology, Software, Visualization, Writing – original draft. YH: Investigation, Supervision, Writing – original draft. WF: Investigation, Supervision, Validation, Writing – review & editing. JL: Supervision, Validation, Writing – original draft. LL: Conceptualization, Formal analysis, Funding acquisition, Supervision, Writing – original draft, Writing – review & editing. TL: Formal analysis, Supervision, Validation, Writing – original draft, Writing – review & editing.