Thyroid nodules can be detected by ultrasound (US) in 19%-67% of the general population. Most detected nodules are benign and without clinical significance, but up to 7%-15% are malignant (1). Among all cancers worldwide, the incidence of thyroid carcinoma ranks 11^th in all patients and sixth in females (2, 3). The incidence of medullary thyroid carcinoma (MTC) represents <5% of all thyroid carcinomas (4, 5).

The risk stratification system of the American College of Radiology Thyroid Imaging Reporting and Data System (ACR-TIRADS) was proposed in 2017 (6). Its main purpose is to assess the risk of malignancy and help decide whether a US-guided fine-needle aspiration (US-FNA) and follow-up should be performed (7, 8). Owing to the mismatch of classification systems used by different hospitals, the TI-RADS is affected by low diagnostic specificity (9) and is difficult to use in everyday clinical work (10, 11). In search of better diagnostic options, local institutions worldwide have issued additional guidelines: Chinese-TIRADS (C-TIRADS) (8), European-TIRADS (EU-TIRADS) (12), and other diagnostically-modified TIRADS (11). Other diagnostic guidelines include the Korean Society of Thyroid Radiology (KSThR) guidelines published in 2017 (13), the American Thyroid Association (ATA) guidelines published in 2015 (14), and the American Association of Clinical Endocrinologists (AACE), the American College of Endocrinology (ACE) and the Associazione Medici Endocrinologi (AME) guidelines published in 2016 (15). In addition, diagnostic imaging has gradually begun to take advantage of artificial intelligence (AI) and machine learning in applying TI-RADS (16–19).

Nevertheless, the existing TIRADS mainly focuses on papillary thyroid carcinoma (PTC), with less attention being paid to MTC. Although MTC accounts for only <5% of thyroid carcinomas, MTC is characterized by a rapid progression and a worse prognosis than PTC, resulting in approximately 13% of all thyroid cancer-related deaths (3, 20). In addition, MTC has a low sensitivity to chemoradiotherapy and a high recurrence rate postoperatively (21). Thus, it is important to identify MTC early to improve patient outcomes. US is the first option for the early screening of thyroid nodules due to its safety, convenience, and good display ability of small lesions (1).

Different treatments and prognoses of MTC and PTC substantiate the need for more precise tools to diagnose them adequately as soon as possible. In addition, selecting the best diagnostic guidelines remains controversial, and applying such guidelines relies on local imaging practices and sonographers’ experience, introducing some subjectivity in the assessment. The AI-SONIC™ artificial intelligence assistant diagnostic system (Zhejiang Demetics Medical Technology Co., Zhejiang, China) is diagnostic tool able to sort US images automatically (22–25); it was developed by deep learning from the US image and pathological data of >200,000 thyroid nodules (26). Still, the comparisons between the AI-SONIC™ and the available guidelines are unknown.

Therefore, the purpose of this study was to analyze the diagnostic value of the six US guidelines and an artificial intelligence system (AI-SONIC™) in differentiating malignant vs. benign thyroid nodules.

Although some differences in performance were observed, all six guidelines and the AI-SONIC™ system had satisfactory value in differentiating benign vs. malignant thyroid nodules.

The risk stratification systems for thyroid nodules (ACR- TIRADS and others) are often affected by a low diagnostic specificity. Recent studies noted the same limitations in applying those guidelines as in the present study (7, 10, 27–32). Significant differences were noted between MTC and PTC features: hyper- or isoechoic nodules more often described in MTC and lobulated margins in PTC. The most notable feature of MTC was the increased blood flow, both inside and around the nodules (Figure 2).

Peng et al. (29), in a study of 230 thyroid nodules in 2020, observed that as much as 19.6% of the malignant nodules (45 of 230) did not match any pattern of the ATA guidelines; hence, the ACR- TIRADS guidelines derived the highest diagnostic performance and the greatest level of sensitivity, compared with the AACE/ACE/AME and ATA guidelines. The present study showed comparable results in the malignant (MTC+PTC) group but not in the MTC group, where the KSThR guidelines (a classification system that Peng et al. did not evaluate) demonstrated the highest specificity and accuracy, followed by AACE/ACE/AME. A recent study by Pandya et al. (28) also examined the performance between the ACR -TIRADS and ATA guidelines and showed that both guidelines had similar diagnostic accuracies but that the ACR- TIRADS guidelines resulted in fewer nodules being recommended for immediate FNA and more nodules being recommended for imaging surveillance. In the present study, the accuracy of the ATA guidelines in assessing the risk of MTC was slightly higher than the ACR-TIRADS guidelines due to the higher specificity. Nevertheless, a higher number of unclassifiable cases in the ATA guidelines might cause problems in clinical application, especially by less experienced sonographers.

The application of the US classification systems has local specifics related to the lexicon and diagnostic traditions (11). Shen et al. (33) compared the ACR -TIRADS, ATA, EU-TIRADS, and KSThR guidelines and noted that all four risk-stratification systems had good diagnostic performances. Zhang et al. (31) compared the ACR- TIRADS, ATA, Kwak TI-RADS, and KSThR guidelines and reported that the ACR TI-RADS and Kwak TI-RADS guidelines had better diagnostic performance than the other guidelines in the malignant group. Zhang et al. (31) also noted that among the suspicious US image features, the most significant independent predictor for malignancy was hypoechogenicity. In the present study and Zhu et al. (11), hypoechogenicity was more associated with PTC than with MTC, while in MTC, hyper- or isoechoic nodules were described more often (20.0% vs. 11.8%). Conversely, due to the notable US differences, Zhu et al. (11) recently proposed a modified TI-RADS system specifically for diagnosing MTC. This modified system warrants further investigation.

Interestingly, in the recent study by Matrone et al. (34), among 152 consecutive patients with MTC, US high risk of MTC malignancy included in the EU-TIRADS, 2015 ATA, AACE/ACE-AME, ACR-TIRADS, and K-TIRADS guidelines, varied from 45.4% to 47.4%, while in the present study, it was notably higher (63.7%-73.8%). Li et al. (27) compared the value of the KWAK TI-RADS and 2015 ATA guidelines in ninety-three patients (29 with MTCs, 31 with PTCs, and 33 with thyroid adenomas). The ATA guidelines showed higher specificity and sensitivity for PTC (67.7% and 77.4%, respectively) than MTC (62.1% and 65.5%, respectively), which is consistent with the present study but also slightly lower: 80.5% and 70.9% for MTC, and 94.4% and 80.8% for PTC. Yun et al. (35) evaluated TI-RADS for MTC and reported that 95% of the nodules were classified as either highly suspicious (68%) or intermediately suspicious (26%), which were lower compared with the present study. The selection process might explain this difference based on the FNA results and the generally smaller nodules in the present study. Markedly, Hahn et al. (36) explored the diagnostic efficacy of the ATA and K-TIRADS guidelines in MTC and noted that the nodule size correlated with the diagnosis, and small MTC nodules were classified more commonly as highly suspicious. Nevertheless, the diagnostic value of the classification systems explored in the present study for MTS based on AUC were ATA (0.784) > ACR (0.772) > KSThR (0.763) > C-TIRADS (0.759) > AI (0.748) > EU-TIRADS (0.723) > AACE/ACE/AME (0.713), which is mostly consistent with the results obtained by the studies mentioned above.

Machine deep learning is another approach gaining popularity among scientists and physicians. The literature shows promise in applying AI technology to avoid unnecessary biopsies (37). In the study by Wang et al. (38), the performance of the YOLOv2 neural network did not significantly differ from that of the radiologists (P > 0.05). The AIBx algorithm, proposed by Johnson et al. (39), demonstrated better sensitivity, specificity, and PPV than the radiologists. AI-assisted TI-RADS improved specificity while maintaining sensitivity (18) and avoiding unnecessary biopsies (30). In the present study, the AI-assisted diagnostic algorithm proposed by AI-SONIC™ (Demetics) was compared with other risk assessment guidelines. The results revealed the highest specificity (85.6%), PPV (67.5%), and accuracy (78.6%) of the AI in the MTC group but average sensitivity (63.4%). In some atypical cases, the assistance of AI might avoid misdiagnosis (Figure 3). Image similarity AI models, as a part of diagnostic algorithms, can decrease subjectivity and increase confidence in the predictions (39). Overall, AI models could become a decision-support tool but are not yet ready to substitute for human physicians. Recently, Kang et al. (40) highlighted the usefulness of computer-aided diagnosis, especially for less experienced radiologists and especially for diagnosing PTC. Still, there are differences in performance among the various AI systems, and future studies should also examine and compare the available AI systems.

MTCs were mainly solid, and cystic changes in MTC nodules were rare (7.5%), which is consistent with the conclusions of Lee et al. (41). In benign nodules, cystic and solid structures were more common, while PTC was also mainly solid, and this feature was more significant than in MTC. MTC is mainly hypoechoic, which is consistent with the conclusions of Saller et al. (42) and Zhu (11, 43), while isoechoic, and hyperechoic nodules were more common in benign nodules, which is consistent with the pathological basis of MTC tumor cells being homogeneous, arranged in solid patches and nests, and with less stroma. PTC was also mainly hypoechoic. MTC is prone to suspicious extrathyroid extension. MTC is often invasive. Only about half (48%) of the lesions were localized, while 35% of the lesions broke through the thyroid capsule and invaded the surrounding tissues or local lymph node metastasis. Because the definition of extrathyroid extension in this study included the adjacent capsule, the occurrence rate of suspected extrathyroid extension was higher than in other similar studies. The risk of extrathyroid extension of malignant nodules was 61%, of which 31% were visible to the naked eye, and the specificity of capsule contiguity was poor. There was a >2-mm normal thyroid parenchyma between the nodules and the continuous capsule, which reduced the risk of extrathyroid extension under the microscope to <6%, and there was little or no chance of gross invasion. The suspected lymph node metastasis in the MTC group was significantly different from benign nodules and more significant than in the PTC group. MTC is a rare disease with no significant malignant features. It is difficult even for experienced sonographers to detect it early, and MTC is highly invasive and prone to lymph node metastasis (24, 42). These features are different from PTC and benign nodules, but the lymph node metastasis of MTC will affect the prognosis of patients. Therefore, it is meaningful to detect MTC lesions as soon as possible through the nodule’s characteristics before metastasis. The blood flow of MTC was abundant. The peripheral blood flow of MTC was like that of benign nodules, while the internal blood flow of MTC was significantly richer than that of benign nodules, while the peripheral and internal blood flow of PTC nodules was mostly absent. Although it differed from PTC, the blood supply of MTC overlapped with that of benign nodules, and the significance of blood flow differentiation was limited. In this study, it was found that coarse calcification was most common in MTC (33.8%), followed by PTC (28.7%), and less common in the benign group (21.7%), but the difference was not statistically significant. Microcalcifications were more common in the PTC group (49.7%), followed by the MTC group (35.0%), while it was rarer in the benign group (18.9%), but when the three groups were compared at the same time, the difference of microcalcification was only statistically significant between the PTC and benign groups. The main reason for calcification formation in MTC is that the local cancer tissue forms a calcium and phosphorus deposition microenvironment under the effect of bone matrix protein regulating cell matrix, and then amyloid surrounds the calcium and phosphorus deposition (44). Therefore, coarse calcification in MTC is slightly more common than in PTC. In this study, microcalcification and coarse calcification are common in the MTC group. Different studies have reported different types of calcifications and MTC, which might be related to the course of MTC nodules (45–47). PTC nodules also have many fibrous stromal hyperplasia, sclerosis, and rich calcareous, and coarse calcification was not rare, but microcalcification was more common in the PTC group. Microcalcification in PTC might be due to poor blood supply, and necrotic calcification occurs easily (48–50). Nevertheless, the benign nodules undergoing surgery may cause compression symptoms to the surrounding tissues and affect the appearance due to the large volume of the nodules, and the nodules have suspicious ultrasonic features, so the calcification rate of benign nodules in this study is also high. On the other hand, the benign nodules encountered in daily work might not be calcified so frequently, and this feature might still have a certain differential value.

Therefore, based on the above, regarding nodule size, it is easy to miss the diagnosis because the ultrasonic features of early MTC are not obvious. Compared with PTC, the volume of MTC is usually larger when it is found. Most studies showed that most MTCs are solid nodules with few cystic changes, while PTCs rarely have cystic changes and benign nodules often have cystic changes (42, 45, 46). Most MTCs showed low echo, followed by very hypoechoic, while hyperechoic was rare. Hypoechoic PTC nodules were also the most common, while hypoechoic, isoechoic, and hyperechoic benign nodules were not uncommon. Most authors agree that MTC nodules have clear boundaries, but some hold opposite opinions; an unclear boundary might be related to malignancy, but it might also be caused by an uneven echo of thyroid parenchyma caused by Hashimoto thyroiditis and other reasons (42, 45, 46). Of note, Hashimoto thyroiditis appears to coexist with differentiated thyroid cancers (51). There are different reports on whether MTC nodules are mainly regular in morphology. Many authors believe that MTC is mostly regular in morphology (round and quasi-round), but some scholars also found that most MTC nodules are irregular in morphology (42, 45, 46). PTC often has irregular margins, and benign nodules usually have regular margins. The aspect ratio of MTC is mainly <1, that of PTC is >1, and that of benign nodules is <1. Studies suggest that the calcification in MTC nodules is mainly amyloid deposits, so the proportion of coarse calcification is higher than that of microcalcification in MTC (42, 45, 46). In PTC, microcalcification is more common due to necrosis foci. MTC is often invasive, PTC can also have extrathyroid extension, but it is less common than MTC, while benign nodules are non-invasive. MTC often shows abundant blood flow signals in and around the nodules. PTC mainly lacks blood supply. Blood flow signals can appear in benign nodules.

The role of contrast-enhanced US (CEUS) is defined in diagnosing and managing thyroid cancer (52). CEUS is particularly useful for revealing the vascularization of tissues and is sensitive and specific for thyroid cancer (52). Still, whether CEUS could be used to differentiate PTC vs. MTC remains to be confirmed. Future studies should include multimodality US.

This study had some limitations. First, because of the low incidence of MTC, the sample size of this was relatively small, which might reduce the power of the study to some extent. Secondly, only MTCs and PTCs were included. Other malignant nodules, such as follicular carcinoma and undifferentiated carcinoma, were not included. Indeed, PTC is the most common subtype of thyroid carcinoma, but MTC has a worse prognosis. In addition, because PTC is more frequent, the classification guidelines are mainly based on the PTC features. Therefore, this study aimed to examine MTC but used PTC as controls. Future studies should also examine other subtypes. Third, only one center was involved. Since US is operator-dependent (40), the generalizability of the conclusions might be limited. Multicenter studies are necessary to address this issue. Finally, because of the retrospective nature of the present study, some US characteristic data were incomplete and could not be included in the analysis. These characteristics can be considered in future prospective studies.

In conclusion, six classification guidelines and the AI-SONIC™ system efficiently differentiate benign vs. malignant thyroid nodules. The C-TIRADS guidelines demonstrated the best performance in the authors’ setting. The ATA and ACR guidelines and the AI-SONIC™ system were effective tools for identifying MTC nodules. Thus, using more than one diagnostic tool could be associated with a more satisfactory differentiation among MTC, PTC, and benign nodules.

The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding author.

The studies involving human participants were reviewed and approved by Fujian Provincial Hospital (K2019-01-051). Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

Conceptualization: LY, NL. Data curation: LY, MW, GC. Formal analysis: LY, MW. Funding acquisition: NL. Investigation: LY, MW, GC. Methodology: LY, NL. Project administration: LY, MW. Resources: LY, NL. Software: LY, MW. Supervision: LY, MW, GC. Validation: LY, MW. Visualization: MW, GC. Writing-original draft: LY. Writing-review & editing: LY. All authors contributed to the article and approved the submitted version.