A narrative review on innovations of thyroid nodule ultrasound diagnosis: applications of robot and artificial intelligence technology

Introduction

With the continuous development of ultrasound medicine, the increase of health examination and the change of environment and lifestyle, the detection rate of thyroid nodules is increasing. Although most thyroid nodules are benign, about 5–15% are still malignant, so early screening and evaluation of thyroid nodules are of great significance (1). However, the number of sonographers is relatively small and the average professional level is limited. How to improve the efficiency of ultrasound in the detection of thyroid nodules, especially malignant nodules, has become an important problem. With the continuous progress of science and technology, the diagnostic methods of thyroid diseases are also evolving. From the initial manual palpation and simple imaging examination to today’s high-resolution ultrasound imaging and computer-aided diagnosis technology, the diagnostic methods of thyroid diseases have undergone significant changes. In recent years, with the continuous innovation of robotic automated scanning and ultrasound image intelligent-assisted diagnosis technology (2,3), the construction of a digital (4) and intelligent medical diagnosis platform for thyroid nodules (5) has become a feasible way for early screening and management of thyroid nodules (Figure 1). This article focuses on the research status of thyroid robot automated scanning and intelligent assisted diagnosis technology, and put forward the application prospect of intelligent ultrasound diagnosis in the examination of thyroid nodules, analyze its advantages and challenges in actual clinical practice, so as to provide reference for the research of intelligent thyroid nodule medical platform. We present this article in accordance with the Narrative Review reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-75/rc).

Figure 1 Thyroid ultrasound intelligent diagnostic process. AI, artificial intelligence; LLMs, large language models.

Methods

In this study, we systematically retrieved the PubMed and Web of Science databases and included the relevant articles on the application of robots and artificial intelligence (AI) in thyroid ultrasound diagnosis published from 2004 to 2025. The types of articles include original research papers, review articles and conference papers. Then we summarized and analyzed the existing studies, as seen in Table 1.

Results

Automated thyroid robotic scanning

The automated thyroid robotic scanning technology provides a new solution for thyroid examination through accurate positioning, moment arm control and trajectory planning (6,7). Thyroid robot effectively reduces human error through automatic operation, improves the standardization of scanning process, and then enhances the accuracy and reliability of clinical diagnosis (8,9). The core of this technology lies in the precise positioning and trajectory tracking of the thyroid. Real-time ultrasound images are used to identify the anatomical structure of the thyroid, extract its boundaries and key features, and match them with the preset scanning trajectory to correct the position deviation and accurately determine its spatial position. Although this method significantly improves the accuracy of scanning, it may still face the problem of insufficient local detail capture when dealing with complex or dynamic anatomical structures. By combining real-time image feedback mechanism and adaptive optimization algorithm technology, the ability to identify thyroid details can be further improved. Zielke et al. (10) control the movement of the robot arm based on the thyroid lobe segmentation and the anatomical structure of ultrasound image feedback, and study the robot guidance algorithm to promote the calculation accuracy of thyroid volume and effectively improve the positioning accuracy. This study compared the volume calculation accuracy of a robotic thyroid scanning system based on a three-dimensional (3D) volume algorithm with manual measurements by a conventional sonographer. The experimental results showed that the average error of the automated robotic measurements was 8.23%, which was significantly lower than the 20.85% of the manual measurements made by the traditional sonographer. Notably, with the assistance of the robotic system, the sonographer’s measurement accuracy was improved by 61%. Accurate calculation of thyroid volume has important clinical guidance value in ablation procedures for malignant thyroid nodules. Firstly, accurate volume measurement can provide a quantitative basis for ablation scope planning. Secondly, it can help preoperative energy dose calculation and surgical plan development. Finally, it can be used as an important reference index for postoperative efficacy assessment. The automated measurement method shows higher accuracy and repeatability than traditional manual measurement, providing more reliable data support for clinical decision-making. However, the positioning stability still needs to be improved in the case of rapid movement or many interference factors. Combining the spatial information acquired by the external camera with the anatomical data of the ultrasound image is expected to further enhance the robustness and accuracy of the system. In terms of mechanical control, the thyroid robot monitors the contact state between the probe and the skin in real time through a force sensor, and dynamically adjusts the angle of the probe to maintain the appropriate contact force and avoid inappropriate pressure affecting the imaging quality. Tsumura et al. (11) used a series elastic drive end-actuator to achieve a stable contact force to ensure high-quality image acquisition. However, it is still a challenge to achieve stable scanning imaging in complex clinical environments in response to patient body size differences, which can be optimized in the future combined with personalized pressure regulation strategies. In terms of trajectory feedback control, the robot dynamically adjusts the scanning path through predefined rules combined with real-time ultrasound data to effectively improve the scanning efficiency. Zhou et al. (12) accurately tracked and positioned the thyroid based on network visual servoing technology and verified the effectiveness of the method. However, the flexibility and adaptability of the system still need to be further improved when dealing with complex or changing scanning targets. In the future, the deep learning models can be used to strengthen the detection and real-time feedback adjustment of thyroid ultrasound images to improve the response ability of the system to abnormal scanning imaging. In addition, the combination of deep learning and reinforcement learning provides a new idea for the optimization of probe trajectory and scanning angle. The deep Q-Network method proposed by Su et al. (13) combines the advantages of deep learning and reinforcement learning to optimize probe trajectory design and improve scanning efficiency and quality. And this research proposed a fully autonomous system for thyroid ultrasound sweeping and intelligent assisted diagnosis, and by comparing the nodule segmentation performance of different AI algorithms, it was found that the improved deep learning algorithm VariaNet+ performs optimally, and its segmentation index reaches the Intersection over Union (IoU) of 0.6369, the Dice of 0.7474, and the precision of 0.8007. At the same time, the study adopted the Bayesian optimization algorithm to improve the sweeping imaging stability, and the experimental results showed that the system can obtain the minimum entropy value at 21.3 s, when the imaging quality is the best. However, the clinical application of this method may be limited by the computational power and the accuracy of real-time feedback, and its real-time performance and reliability can be further optimized by more efficient algorithms and hardware in the future. With the breakthrough of 5G communication technology, Zhang et al. (14) developed a remote thyroid ultrasound examination robot based on 5G, which realized remote thyroid scanning and diagnosis, and opened up a new path for improving the accessibility and diagnostic efficiency of medical services. Although thyroid robotics has made significant progress in improving the standardization and efficiency of ultrasound examination, some studies have suggested that participants still have some doubts about the safety and comfort of the robotic scanning process. The main reason is that robots lack clinical experience and medical knowledge accumulation, and cannot provide sufficient patient trust and reassurance like experienced doctors (13). Therefore, future research should focus on strengthening the intelligence level of robots in the clinical environment, and combine the judgment and feedback of medical experts to improve the adaptability of robot operation and patient acceptance. Through continuous technical improvement and clinical validation, thyroid robotic automated scanning technology is expected to provide a more reliable and efficient solution for the early diagnosis and personalized treatment of thyroid diseases (15).

With the combination of thyroid ultrasound robot system and intelligent auxiliary diagnosis, the intelligent thyroid medical diagnosis system integrating scanning and diagnosis can quickly provide accurate diagnosis of benign and malignant thyroid nodules based on the analysis of large sample case data. In the context of uneven distribution of medical resources, fully automated thyroid robots are expected to solve the problem of lack of professional health services in rural and remote areas. However, the flexibility of thyroid robots is currently not comparable to that of experienced sonographers. In a complex and variable clinical environment, thyroid robots mainly rely on preset programs to perform scanning tasks, which are difficult to adjust in time according to actual clinical mutations. Therefore, it is of great significance to develop an intelligent thyroid robot system that can flexibly cope with diverse clinical environments and accumulate diagnostic knowledge from practice like a clinical sonographer.

Intelligent diagnosis of thyroid nodule ultrasound images

The diagnostic performance of traditional ultrasound is highly dependent on the clinical experience of the physician, which often leads to strong subjectivity and low inter-observer variability in diagnostic results. This limitation is especially significant in the clinical practice of patients in areas with relatively scarce medical resources or with junior doctors. In recent years, the application of AI technology in the field of ultrasound image diagnosis of thyroid nodules provides new ideas for solving the above problems. The intelligent auxiliary diagnosis system based on deep learning can realize the automatic extraction and accurate classification of lesion features, and significantly improve the objectivity and repeatability of diagnosis.

Rule-based feature matching

In the early stage of the research on intelligent diagnosis of thyroid nodules ultrasound images, researchers mainly used rule-based feature matching and artificial feature engineering methods. Clinical experts manually define morphological features such as boundary clarity and internal echo of nodules, and establish the corresponding diagnostic rule base for pattern matching (16-18). The diagnostic process of this method is transparent, and the rules are interpretable, so that doctors can intuitively understand the diagnostic basis. However, its high reliance on expert experience may miss important feature information, which is difficult to effectively adapt to complex and variable case data. Meanwhile, artificial feature extraction is easily affected by subjective factors, resulting in poor reproducibility of diagnostic results. To overcome these limitations, researchers have gradually turned to methods based on quantitative feature analysis to improve the objectivity and stability of diagnosis. Gul et al. (19) proposed a statistical feature quantification method, and constructed a diagnostic model based on quantitative indicators by calculating morphological parameters such as nodule shape index, edge regularity and boundary clarity. Compared with the traditional rule matching method, this method enhances the objectivity of diagnostic features and reduces subjective intervention. However, the feature selection process of this method still relies on expert experience, which may lead to the omission of potentially diagnostic relevant features, and it is difficult to avoid the selection bias artificially introduced. With the breakthrough of deep learning technology, researchers have begun to explore intelligent diagnosis methods based on automatic feature learning. A large number of studies have shown that intelligent diagnosis methods based on deep learning not only overcome the limitations of traditional artificial feature engineering, but also achieve end-to-end automated diagnosis on a data-driven basis (20-22). Although the earlier rule matching and artificial feature extraction methods have many limitations, they have laid an important theoretical and practical foundation in the development of intelligent diagnosis technology for thyroid nodules.

Intelligent segmentation and classification of thyroid nodule ultrasound images

In recent years, deep learning technology has made significant progress in the field of ultrasound image segmentation and classification of thyroid nodules. The automatic feature extraction method based on deep neural network effectively overcomes the limitations of traditional manual feature extraction methods in processing complex texture and irregular shape medical images through multi-level feature learning mechanism.

Accurate segmentation of thyroid nodules is a key research direction for benign and malignant judgment and lesion analysis, which provides a reliable basis for intelligent diagnosis. U-Net is a classical neural network architecture for thyroid nodule ultrasound image segmentation. Its encode-decoder structure can effectively capture the multi-scale features of the nodule, and maintain spatial information through jump connection to realize the stepwise processing and reconstruction of thyroid nodule ultrasound images. Etehadtavakol et al. (23) combined U-Net and VGG16 to improve the segmentation accuracy by using transfer learning under the condition of limited data sets. However, U-Net still has the problem of limited feature expression ability, which makes it difficult to make full use of deep information. To this end, researchers proposed U-Net++ to enhance the feature transfer ability through dense jump connections. Nie et al. (24) used U-Net as the backbone network to obtain rich nodule feature representation through dense full convolutional neural network (CNN), and achieved a Dice coefficient of 0.9367, which significantly improved the segmentation accuracy. To further optimize the segmentation ability of the model, attention mechanisms were introduced into the segmentation network. SK-Unet ++ proposed by Dai et al. (25) used the selective kernel attention mechanism to adjust the receptive field, optimize the capture of multi-scale spatial features, and improve the adaptive segmentation ability of thyroid nodules. Gong et al. (26) adopted the pilot attention method and introduced prior knowledge to guide the accurate segmentation of nodule regions. Wang et al. (27) innovatively used the dual-path attention mechanism, in which one channel captured the global information through the lightweight cross-channel interaction mechanism, and the other channel focused on the edge and surrounding information of nodes through the residual bridge network to achieve accurate edge segmentation. However, while the attention mechanism improves the performance of the AI model, it also increases the training complexity and may lead to overfitting due to insufficient data. Therefore, Yang et al. (28) developed a lightweight U-shaped network combined with efficient convolution and attention mechanism to reduce computational complexity and improve the generalization ability of the model while ensuring the segmentation accuracy. In view of the challenges of variable morphology and fuzzy boundary of thyroid nodules, multi-scale feature fusion technology has become a key strategy to improve the segmentation performance. Mi et al. (29) proposed a pyramid pooling module to integrate feature information of different scales to adapt to nodules of different sizes and shapes. This method optimizes the 3D visualization of thyroid ultrasound images and improves the ability to distinguish thyroid nodules from surrounding tissues. However, due to the limitation of ultrasound video data, its effect in clinical application still needs to be further verified. Future researches can make the models more generalized, optimize the utilization of computational resources, and improve the clinical feasibility.

The classification of benign and malignant ultrasound images of thyroid nodules is the core link of intelligent diagnosis, which can assist clinical decision-making, improve the accuracy of diagnosis, and thus reduce unnecessary biopsy and surgery. In the task of benign and malignant classification of thyroid nodules, CNN is widely used because of its excellent image feature extraction ability. Chen et al. (30) conducted a multicenter study on the classification of benign and malignant thyroid nodules by CNN and confirmed the effectiveness of AI-assisted diagnosis in reducing unnecessary fine needle aspiration (FNA). With the improvement of computing power and data size, researchers improve the classification accuracy by increasing the network depth. Zhao et al. (31) used deep convolutional neural network (DCNN) to improve the multi-category recognition ability of thyroid diseases and enhance the sensitivity and specificity of the model. However, deep neural networks may have a gradient vanishing problem, which affects the training stability. Therefore, the residual network (ResNet) is introduced into the classification task to maintain the gradient flow and improve the feature representation ability. Hang et al. (32) combined residual generative adversarial network (GAN) and random forest classifier to achieve a classification accuracy of 0.95, which further improved the performance of benign and malignant classification of thyroid nodules. In the classification task, the attention mechanism can also help the model to focus on the key feature regions and improve the classification accuracy. Lu et al. (33) used nodule mask-guided classification model combined with deformable attention network to extract more fine-grained features, which effectively improved the performance of benign and malignant classification. However, the introduction of attention mechanisms also increases the computational burden of the models and may reduce the real-time performance of practical applications. Therefore, future researches can combine lightweight attention modules to optimize computing resource allocation and make it more suitable for clinical applications.

In recent years, Transformer architecture has emerged in the field of medical image analysis with its self-attention mechanism and ability to efficiently capture global feature information. Vision Transformer (ViT) can pay attention to the nodule morphology globally and improve the boundary recognition ability, while Swin Transformer realizes local-global feature extraction through the hierarchical window attention mechanism. Huang et al. (34) combined the residual block with Swin-Transformer to improve the sensitivity of local and global features of nodules and achieve a classification accuracy of 0.8832. However, the Transformer models relies on large-scale labeled data for training and consumes large computational resources, making its efficient deployment on medical devices a challenge. Future research directions are knowledge distillation or transfer learning methods to reduce the dependence on large-scale data, while optimizing the model architecture to reduce computational cost and improve clinical usability.

Multimodal diagnosis of thyroid nodule ultrasound images

Gray-scale ultrasound images of thyroid nodules are the most commonly used thyroid imaging method, which can provide key information such as morphology, boundary, echo characteristics and calcification of nodules, and play an important role in the assisted diagnosis of deep learning technology. However, the morphological features of single gray-scale ultrasound image can’t fully reflect the biological characteristics of nodules, which affects the accuracy of the diagnosis of benign and malignant nodules. Therefore, researchers have begun to explore multimodal image fusion methods to improve the accuracy and reliability of intelligent diagnosis. Multimodal diagnostic methods integrate different types of medical image data to provide more comprehensive information of nodules and enhance the segmentation and classification ability of the models. Xiang et al. (35) input gray-scale ultrasound, shear wave elastography and doppler ultrasound images into the pre-trained ResNet18 network for feature extraction, and use the multimodal multi-head attention mechanism to fuse common features to achieve more accurate ultrasound imaging diagnosis. This study validates the effectiveness of the multimodal approach and provides a rapid and efficient diagnostic tool for sonographers. In addition to ultrasound images, other medical images also provide important supplementary information in the diagnosis of thyroid nodules. Miao et al. (36) used ultrasound images combined with computed tomography images to achieve multimodal prediction of central lymph node metastasis of papillary thyroid carcinoma, which significantly improved the diagnostic performance. Zhang et al. (37) further proposed a multi-modal diagnostic method based on ultrasound and infrared thermal imaging, using an adaptive multi-modal hybrid strategy to weight different modal features, and achieving deep fusion through cross-modal encoder, which has significant advantages in the task of benign and malignant classification.

Although the multimodal fusion method improves the performance of the models, it still faces the problems of inconsistent information between modes and high computational complexity. Future studies can optimize the cross-modality alignment technique and explore lightweight feature fusion strategies to reduce computational overhead and improve the clinical applicability of the models.

Processing of CEUS thyroid video by deep learning method

Compared with static ultrasound images, which mainly rely on morphological features for the diagnosis of thyroid nodules, dynamic images provide temporal dimension information, especially the dynamic blood perfusion characteristics in CEUS video, which is crucial for the diagnosis of benign and malignant nodules. The deep learning method based on CEUS can more accurately analyze the microvascular characteristics of lesions by extracting blood perfusion patterns, and improve the reliability of diagnosis. Wan et al. (38) introduced perfusion excitation gate and cross-attention time aggregation module to enhance key features and perform global aggregation, so as to deal with the temporal features of CEUS. Through the uncertainty estimation strategy, the model can accurately locate the key enhancement points, realize video-level nodule segmentation, and perform real-time observation and lesion characterization of microvascular perfusion, which provides a new idea for the intelligent diagnosis of thyroid nodules. However, this model has poor adaptability when dealing with heterogeneous enhancement sequences, resulting in limited generalization ability to different types of lesions. The adaptive temporal attention mechanism can be combined to improve the adaptability of the models to different perfusion modes. Chen et al. (39) proposed a method based on dynamic Swin-Transformer encoder and multi-level feature collaborative learning, which was embedded into the U-Net model to achieve accurate representation of CEUS long-distance enhanced perfusion patterns, effectively segment lesions with fuzzy boundaries, and improve the performance of nodule segmentation. However, the model still has room for optimization in selecting the key frames of CEUS video sequences. Improving the keyframe selection strategy can not only improve the computational efficiency, but also further improve the diagnostic accuracy. In addition, the proposed model has only been trained and validated on a single-center dataset and is limited to the CEUS dataset, so its generalization ability still needs to be further improved. CEUS video has important research value in the intelligent diagnosis of thyroid nodules. The development of more adaptive time series modeling methods, optimization of key frame selection strategies and multi-center research are of great significance for the clinical application of intelligent diagnosis of thyroid nodules.

Interpretability of intelligent diagnosis of thyroid nodule ultrasound images

Despite the continuous development of deep learning technology, its performance in the diagnosis of thyroid nodules continues to improve, but the interpretability of models is still an important challenge that limits its clinical application. At present, deep learning models mainly extract features from raw ultrasound images and perform classification or regression based on these features. However, the decision-making process of these models is usually a non-transparent “black box” mechanism, which makes it difficult for doctors to understand the decision-making basis of these models, thus affecting clinical trust. Therefore, improving the interpretability of the models is crucial to promote its application in medical image analysis. In order to enhance the interpretability of deep learning models, multi-task learning methods combining medical prior knowledge have been studied. Zheng et al. (40) proposed a “segmentation + classification” joint modeling method for the diagnosis of thyroid nodules. This model adopts a multi-scale segmentation network for accurate nodule segmentation, and designs a three-branch classification network combined with expert knowledge to extract features from the original image, regional image and edge image, and obtain classification features through cross-level feature fusion strategy. This method effectively referred to the clinical diagnosis logic of doctors, and improved the ability of the model to distinguish benign and malignant nodules. However, when the dataset is small or the class distribution is uneven, the performance of the model may decrease and it is difficult to handle complex scenarios or specific types of nodules. For further optimization, data augmentation techniques and other methods can be introduced to improve the adaptability of the models in different cases. Another strategy to improve the interpretability of the models is to use feature visualization methods to make the decision-making process of the models more transparent. Liu et al. (41) used Gradient-weighted Class Activation Mapping (Grad-CAM) technology to visualize the key areas of concern of the model in the prediction process, and generated heat maps to highlight features such as calcification, solid echo and high echo intensity, so as to visually show the lesion areas of concern in the AI diagnosis process. This approach enhanced the visual interpretation of model decisions and enables physicians to understand the diagnostic logic of AI models more intuitively. In addition to feature visualization, knowledge distillation is also an important means to improve the interpretability of models. Li et al. (42) used deformable attention networks and distillation-driven interaction modules to extract discriminative features of nodules and provide interpretable predictions at feature locations. Experimental results show that the proposed method can effectively improve the diagnostic performance on large-scale datasets and provide transparent decision-making basis for clinical practice. However, this study only trained based on gray-scale ultrasound images and did not make full use of multimodal information. In addition, this method still has some limitations in identifying nodules with fuzzy boundaries. Wang et al. (43) proposed a multi-granular hierarchical label learning method based on region and boundary knowledge, and combined with Thyroid Imaging Reporting and Data System (TI-RADS) to build interpretable deep learning networks. This method made use of medical domain knowledge to improve the interpretability of the model on specific features. However, the model was mainly trained on a single-center dataset, and the generalization ability still needs to be further validated. Improving the interpretability of deep learning models is crucial for the intelligent diagnosis of thyroid nodules. Future research can focus on combining medical knowledge with multimodal information, optimizing visualization methods, and using knowledge distillation, hierarchical labeling and other methods to enhance the transparency of model decision-making, and improve the generalization performance of models through multi-center dataset training. To promote the clinical application of deep learning models in the intelligent diagnosis of thyroid nodules.

Application of large language models (LLMs) in thyroid nodule diagnosis

With the continuous development of computer technology, LLMs technology has shown great potential in the intelligent diagnosis of thyroid nodules. These models can integrate multimodal information such as ultrasound images and clinical data to provide more comprehensive diagnostic support and help doctors make more comprehensive and accurate diagnostic decisions. Wu et al. (44) used three big oracle models of ChatGPT 3.5, ChatGPT 4.0 and Google’s Bard to conduct malignant tumor diagnosis research, and the experimental results showed that the LLMs had potential in assisting human doctors in diagnosis. Although LLMs show significant advantages in the intelligent diagnosis of thyroid nodules, because they are trained on large-scale datasets, they may have poor performance in dealing with rare cases, thereby affecting the accuracy of the diagnosis results.

Discussion

Although the field of intelligent assisted diagnosis of thyroid nodules has developed rapidly, its research paradigm still faces many challenges. The training of existing models is highly dependent on the quality of data sets, while different doctors have different judgment criteria for key indicators such as nodule boundary and echo features, resulting in strong subjectivity of data labels. In addition, most studies are based on retrospective data from a single center with a small sample size, and data imbalance leads to poor model performance on rare types of nodules, such as follicular carcinoma. Although deep-learning models perform well in classification tasks, the lack of transparency in their decision-making process reduces clinicians’ trust in AI-assisted diagnosis. Many studies only rely on the ultrasound image itself, ignoring multi-dimensional information such as pathology and genomics, resulting in insufficient discrimination ability in dealing with complex cases, such as papillary microcarcinoma and benign hyperplastic nodules, and unable to effectively evaluate the stability of the AI system in the real clinical scenario. Although current AI-assisted diagnostic systems have demonstrated high diagnostic accuracy in identifying benign and malignant thyroid nodules, the “black box” nature of the decision-making process prevents the system from quantifying, in a clinically comprehensible manner, the contribution of key imaging features, such as nodule echogenicity, margin characteristics, and type of calcification, to the final grading of the diagnosis. There is a clear disconnect between this lack of interpretability and the actual clinical diagnostic process. For example, in the diagnostic decision-making for Bethesda class III nodules, the sonographer needs to explicitly assess the weighted combinations between morphologic features and give FNA recommendations based on the combination of different features (45,46).

Future studies should focus on the strategy of multi-center datasets, which can achieve the optimization of the models performance in the following three aspects. Firstly, the inclusion of nodal data encompassing various pathological types—such as papillary, follicular, medullary, and other rare carcinoma subtypes—significantly enhanced the model’s detection rate across all nodule types. Secondly, integrating scan images from physicians of different seniority effectively reduces the model’s dependence on specific scanning practices, so that the AI system maintains stable diagnostic performance in scenarios used by physicians of different experience levels. Finally, by incorporating imaging data from multiple medical institutions and multiple brands of ultrasound equipment, it not only reduces the risk of model overfitting, but also significantly improves its equipment migration capability. At the same time, new techniques such as semi-supervised learning should be developed to reduce the dependence on fully labeled data and use unlabeled data to enhance the robustness of the models. Research on interpretable AI should be strengthened to provide biological explanations for diagnostic results and improve the credibility of decision-making. In addition, integrating ultrasound images, pathological slides, genomic data and electronic health records to build a cross-modal diagnostic framework can better deal with complex cases.

Conclusions

Ultrasound intelligence-assisted diagnosis of thyroid nodules has great potential, but its development is still restricted by data quality, model interpretability and insufficient clinical validation. Future research should jump out of the framework of a single technology optimization and build a four-in-one research paradigm of “data-algorithm-clinical-ethics”. Through the collaborative innovation of standardized data ecology, interpretable multimodal models, dynamic adaptive technology and ethical policies, we can improve the diagnostic accuracy, promote the intelligent ultrasound diagnosis of thyroid nodules from the laboratory to the real world, and finally realize the closed loop of “patient-centered” precision medicine.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-75/rc

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-2025-75/prf

Funding: This work was supported by the Capital Health Development Research Project (No. Initial 2024-2-4068) and the Research of Ultrasonographic Intelligent Diagnosis of Thyroid Nodules Based on Deep Learning and Transfer Learning.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-2025-75/coif). All authors report that this work was supported by the Capital Health Development Research Project (No. Initial 2024-2-4068) and the research of ultrasonographic Intelligent diagnosis of thyroid nodules based on deep learning and transfer learning. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Grussendorf M, Ruschenburg I, Brabant G. Malignancy rates in thyroid nodules: a long-term cohort study of 17,592 patients. Eur Thyroid J 2022;11:e220027. [Crossref] [PubMed]
2. Verburg F, Reiners C. Sonographic diagnosis of thyroid cancer with support of AI. Nat Rev Endocrinol 2019;15:319-21. [Crossref] [PubMed]
3. Zuo S, Hughes M, Yang GZ. Novel Balloon Surface Scanning Device for Intraoperative Breast Endomicroscopy. Ann Biomed Eng 2016;44:2313-26. [Crossref] [PubMed]
4. Mathews SC, McShea MJ, Hanley CL, et al. Digital health: a path to validation. NPJ Digit Med 2019;2:38. [Crossref] [PubMed]
5. Li LR, Du B, Liu HQ, et al. Artificial Intelligence for Personalized Medicine in Thyroid Cancer: Current Status and Future Perspectives. Front Oncol 2020;10:604051. [Crossref] [PubMed]
6. Lee D, Yu HW, Kim S, et al. Vision-based tracking system for augmented reality to localize recurrent laryngeal nerve during robotic thyroid surgery. Sci Rep 2020;10:8437. [Crossref] [PubMed]
7. Hidalgo EM, Wright L, Isaksson M, et al. Current Applications of Robot-Assisted Ultrasound Examination. JACC Cardiovasc Imaging 2023;16:239-47. [Crossref] [PubMed]
8. Cohen O, Blank A, Meiersdorf S, et al. Impact of high-quality ultrasound following community ultrasound on surgical planning and active surveillance in patients with thyroid cancer. Clin Endocrinol (Oxf) 2021;94:990-7. [Crossref] [PubMed]
9. Kaminski JT, Rafatzand K, Zhang HK. Feasibility of Robot-Assisted Ultrasound Imaging with Force Feedback for Assessment of Thyroid Diseases. Proc SPIE Int Soc Opt Eng 2020;11315:113151D.
10. Zielke J, Eilers C, Busam B, et al. RSV: Robotic sonography for thyroid volumetry. IEEE Robot Autom Lett 2022;7:3342-8.
11. Tsumura R, Tomioka T, Koseki Y, et al. Safe Contact Force Generation for Robotic Thyroid Ultrasound Imaging. IEEE Robot Autom Lett 2024;9:1700-7.
12. Zhou J, Tian H, Wang W, et al. Fully automated thyroid ultrasound screening utilizing multi-modality image and anatomical prior. Biomed Signal Process Control 2024;87:105430.
13. Su K, Liu J, Ren X, et al. A fully autonomous robotic ultrasound system for thyroid scanning. Nat Commun 2024;15:4004. [Crossref] [PubMed]
14. Zhang YQ, Yin HH, He T, et al. Clinical application of a 5G-based telerobotic ultrasound system for thyroid examination on a rural island: a prospective study. Endocrine 2022;76:620-34. [Crossref] [PubMed]
15. Kojcev R, Khakzar A, Fuerst B, et al. On the reproducibility of expert-operated and robotic ultrasound acquisitions. Int J Comput Assist Radiol Surg 2017;12:1003-11. [Crossref] [PubMed]
16. Alexander LF, Patel NJ, Caserta MP, et al. Thyroid Ultrasound: Diffuse and Nodular Disease. Radiol Clin North Am 2020;58:1041-57. [Crossref] [PubMed]
17. Fang M, Lei M, Chen X, et al. Radiomics-based ultrasound models for thyroid nodule differentiation in Hashimoto's thyroiditis. Front Endocrinol (Lausanne) 2023;14:1267886. [Crossref] [PubMed]
18. Ren JY, Lin JJ, Lv WZ, et al. A Comparative Study of Two Radiomics-Based Blood Flow Modes with Thyroid Imaging Reporting and Data System in Predicting Malignancy of Thyroid Nodules and Reducing Unnecessary Fine-Needle Aspiration Rate. Acad Radiol 2024;31:2739-52. [Crossref] [PubMed]
19. Gul M, Bonjoc KC, Gorlin D, et al. Diagnostic Utility of Radiomics in Thyroid and Head and Neck Cancers. Front Oncol 2021;11:639326. [Crossref] [PubMed]
20. Peng S, Liu Y, Lv W, et al. Deep learning-based artificial intelligence model to assist thyroid nodule diagnosis and management: a multicentre diagnostic study. Lancet Digit Health 2021;3:e250-9. [Crossref] [PubMed]
21. Xu J, Xu HL, Cao YN, et al. The performance of deep learning on thyroid nodule imaging predicts thyroid cancer: A systematic review and meta-analysis of epidemiological studies with independent external test sets. Diabetes Metab Syndr 2023;17:102891. [Crossref] [PubMed]
22. Chen C, Liu Y, Yao J, et al. Leveraging deep learning to identify calcification and colloid in thyroid nodules. Heliyon 2023;9:e19066. [Crossref] [PubMed]
23. Etehadtavakol M, Etehadtavakol M, Ng EYK. Enhanced thyroid nodule segmentation through U-Net and VGG16 fusion with feature engineering: A comprehensive study. Comput Methods Programs Biomed 2024;251:108209. [Crossref] [PubMed]
24. Nie X, Zhou X, Tong T, et al. N-Net: A novel dense fully convolutional neural network for thyroid nodule segmentation. Front Neurosci 2022;16:872601. [Crossref] [PubMed]
25. Dai H, Xie W, Xia E. SK-Unet++: An improved Unet++ network with adaptive receptive fields for automatic segmentation of ultrasound thyroid nodule images. Med Phys 2024;51:1798-811. [Crossref] [PubMed]
26. Gong H, Chen J, Chen G, et al. Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules. Comput Biol Med 2023;155:106389. [Crossref] [PubMed]
27. Wang S, Li Z, Liao L, et al. DPAM-PSPNet: ultrasonic image segmentation of thyroid nodule based on dual-path attention mechanism. Phys Med Biol 2023;
28. Yang Y, Huang H, Shao Y, et al. DAC-Net: A light-weight U-shaped network based efficient convolution and attention for thyroid nodule segmentation. Comput Biol Med 2024;180:108972. [Crossref] [PubMed]
29. Mi J, Wang R, Feng Q, et al. Three-dimensional visualization of thyroid ultrasound images based on multi-scale features fusion and hierarchical attention. Biomed Eng Online 2024;23:31. [Crossref] [PubMed]
30. Chen C, Jiang Y, Yao J, et al. Deep learning to assist composition classification and thyroid solid nodule diagnosis: a multicenter diagnostic study. Eur Radiol 2024;34:2323-33. [Crossref] [PubMed]
31. Zhao H, Zheng C, Zhang H, et al. Diagnosis of thyroid disease using deep convolutional neural network models applied to thyroid scintigraphy images: a multicenter study. Front Endocrinol (Lausanne) 2023;14:1224191. [Crossref] [PubMed]
32. Hang Y. Thyroid Nodule Classification in Ultrasound Images by Fusion of Conventional Features and Res-GAN Deep Features. J Healthc Eng 2021;2021:9917538. [Crossref] [PubMed]
33. Lu J, Ouyang X, Shen X, et al. GAN-Guided Deformable Attention Network for Identifying Thyroid Nodules in Ultrasound Images. IEEE J Biomed Health Inform 2022;26:1582-90. [Crossref] [PubMed]
34. Huang L, Xu Y, Wang S, et al. SRT: Swin-residual transformer for benign and malignant nodules classification in thyroid ultrasound images. Med Eng Phys 2024;124:104101. [Crossref] [PubMed]
35. Xiang Z, Zhuo Q, Zhao C, et al. Self-supervised multi-modal fusion network for multi-modal thyroid ultrasound image diagnosis. Comput Biol Med 2022;150:106164. [Crossref] [PubMed]
36. Miao S, Xuan Q, Huang W, et al. Multi-region nomogram for predicting central lymph node metastasis in papillary thyroid carcinoma using multimodal imaging: A multicenter study. Comput Methods Programs Biomed 2025;261:108608. [Crossref] [PubMed]
37. Zhang N, Liu J, Jin Y, et al. An adaptive multi-modal hybrid model for classifying thyroid nodules by combining ultrasound and infrared thermal images. BMC Bioinformatics 2023;24:315. [Crossref] [PubMed]
38. Wan P, Xue H, Liu C, et al. Dynamic Perfusion Representation and Aggregation Network for Nodule Segmentation Using Contrast-Enhanced US. IEEE J Biomed Health Inform 2023;27:3431-42. [Crossref] [PubMed]
39. Chen F, Han H, Wan P, et al. Joint Segmentation and Differential Diagnosis of Thyroid Nodule in Contrast-Enhanced Ultrasound Images. IEEE Trans Biomed Eng 2023;70:2722-32. [Crossref] [PubMed]
40. Zheng Z, Liang E, Zhang Y, et al. A segmentation-based algorithm for classification of benign and malignancy Thyroid nodules with multi-feature information. Biomed Eng Lett 2024;14:785-800. [Crossref] [PubMed]
41. Liu Y, Feng Y, Qian L, et al. Deep learning diagnostic performance and visual insights in differentiating benign and malignant thyroid nodules on ultrasound images. Exp Biol Med (Maywood) 2023;248:2538-46. [Crossref] [PubMed]
42. Li M, Zhou H, Li X, et al. SDA-Net: Self-distillation driven deformable attentive aggregation network for thyroid nodule identification in ultrasound images. Artif Intell Med 2023;146:102699. [Crossref] [PubMed]
43. Wang M, Chen C, Xu Z, et al. An interpretable two-branch bi-coordinate network based on multi-grained domain knowledge for classification of thyroid nodules in ultrasound images. Med Image Anal 2024;97:103255. [Crossref] [PubMed]
44. Wu SH, Tong WJ, Li MD, et al. Collaborative Enhancement of Consistency and Accuracy in US Diagnosis of Thyroid Nodules Using Large Language Models. Radiology 2024;310:e232255. [Crossref] [PubMed]
45. Orlando G, Graceffa G, Mazzola S, et al. The Role of "Critical" Ultrasound Reassessment in the Decision-Making of Bethesda III Thyroid Nodules. Medicina (Kaunas) 2023;59:1484. [Crossref] [PubMed]
46. Scerrino G, Cocorullo G, Mazzola S, et al. Improving Diagnostic Performance for Thyroid Nodules Classified as Bethesda Category III or IV: How and by Whom Ultrasonography Should be Performed. J Surg Res 2021;262:203-11. [Crossref] [PubMed]