Predicting the CLN status is important for guiding the clinical management of patients with PTC. In this study, we integrated the radiomic features of conventional thyroid US images with the gene expression profile to identify potential radiogenomic biomarkers for PTC. We successfully identified 9 significant pairwise associations between radiomic features and gene modules annotated by functional gene enrichment analysis. These associations demonstrated the feasibility of the noninvasive molecular characterization of PTC by using radiogenomic methods, which can provide complementary information for the noninvasive classification and management of PTC.
High-resolution US is regarded as the first choice for the preoperative assessment of the CLN status. However, the efficiency of US in detecting nonpalpable CLNM is unacceptable. In particular, metastatic nodes in the central cervical compartment are not easily detected by US examination because they are obscured by the thyroid gland and complex structures in the central neck. In our study, diagnosing CLNM based solely on conventional US yielded an AUC of only 0.586, with relatively low ACC and SEN. Radiomics is a new technique that converts medical images into high-dimensional mineable data by quantitively extracting high-throughput features [13, 19]. Radiomics sometimes applies machine learning methods and has attracted the research interest of many scholars who seek to explore the association between diagnostic information and imaging features . The AUC of our radiomic signature was improved to 0.873, which indicated that compared with the model relying solely on the US features of LNs, the radiomic model had better performance in predicting CLNM, which is consistent with our previous work .
The radiomic feature ‘Rectlike’ measures the smoothness of the tumour margin, and a higher ‘Rectlike’ value indicates a smoother margin. An ill-defined PTC margin was reported to be significantly associated with CLNM [7, 22]; therefore, patients with a higher ‘Rectlike’ value tended to have a lower risk of CLNM. ‘Rectlike’ was negatively correlated with the modules MEpink and MEmagenta. The gene module MEpink was enriched in biological processes, cellular components, and molecular functions and pathways relating to cell division, while the module MEmagenta represented the upregulation of telomere maintenance via telomerase and cell-cell adhesion. Cancer progression is achieved by uncontrolled cell division, invasion, and, eventually, metastasis . Many anti-cancer agents have been designed to regulate cell division to curb cancer progression and metastasis . Cell division causes the telomere to shorten gradually; hence, a key event in the acquisition of cellular immortality is upregulation of the telomere maintenance mechanism . Telomerase is an enzyme responsible for telomere maintenance. An important mechanism of telomerase activation includes mutations in the promoter region (C228T and C250T) of the telomere reverse transcriptase (TERT) gene . Clinically, mutations in the TERT promoter are frequently examined from FNA samples . PTC tumours harbouring TERT mutations showed more frequent regional LNM spread than did PTC tumours with a wild-type TERT promoter , and TERT mutations are markers for metastatic behaviour .
The shape feature ‘Eccentricity’ measures the ratio of the longitudinal axis to the horizontal axis of the tumour. Tumours with a higher ‘Eccentricity’ value tend to be taller than wide in shape. Previous studies have demonstrated that a taller-than-wide shape on US is an independent predictor for the absence of CLNM [30, 31]. ‘Eccentricity’ was negatively correlated with the module MEsalmon, which was related to cell-cell adhesion and extracellular matrix (ECM) in the GO and KEGG enrichment analyses. Cell-cell adhesion is essential for cell-cell cooperation, multicellular polarity, tissue homeostasis, and collective cell movement . Cancer cell migration, the basis for metastatic dissemination, is a plastic and adaptive process integrating cell-cell adhesion, cytoskeletal dynamics, and ECM remodelling. In single-cell migration, the loss of adhesion between cells triggers a dynamic change in the actin cytoskeleton, which endows the cells with motility and alters cell polarity to form spindle-shaped cells. These newly formed mesenchymal-like cells invade the basal ECM and migrate to the underlying tissues . Degradation and remodelling of the ECM, including the basement membrane, by tumour-secreted proteolytic enzymes are also crucial steps in the process of cancer cell intra- and extravasation and colonization at distant sites .
The position feature ‘Distance to capsule’ indicates the distance between the nodule and the nearest thyroid capsule. Evidence has shown that the tumour being in close proximity to or within the capsule was significantly more indicative of CLNM in PTC [22, 35]. Closer proximity to the thyroid capsule may offer more chances for tumours to encounter lymphatic vessels, therefore increasing the likelihood of metastasis within the lymphatic system . ‘Distance to capsule’ was also correlated with MEsalmon.
Both ‘mean tumour contrast’ and ‘deviation ratio of tumour and normal thyroid gland’ are features that reflect the internal echogenicity of PTC. Tumours with higher ‘mean tumour contrast’ and ‘deviation ratio of tumour and normal thyroid gland’ values tend to be more hypoechoic or even markedly hypoechoic. Lee et al.  demonstrated that hypoechoic and markedly hypoechoic tumours on US were independent risk factors for CLNM, and these malignant US appearances suggested a more invasive biological behaviour, including CLNM . ‘Mean tumour contrast’ was correlated with the module MEsalmon as well, and ‘deviation ratio of tumour and normal thyroid gland’ was correlated with the modules MEpink and MEmagenta.
‘Minimum calcification area’ is a feature that measures the extent of microcalcification of PTC. The higher the ‘minimum calcification area’ value, the more microcalcifications there are within the tumour. Microcalcification has been recognized as an independent predictive factor for CLNM . ‘Minimum calcification area’ was positively correlated with the modules MElightcyan and MEblue. Upon conducting GO and KEGG enrichment analyses, we found that the module MElightcyan was related to complement activation, while MEblue represented cell adhesion and glycolysis. Accumulating evidence has shown that complement activation in the tumour microenvironment promotes tumour growth, suppresses antitumour immunity, and increases metastasis . Enhanced glycolysis has been considered to be the dominant metabolic alteration in malignant tumours  and the primary source of ATP for tumour survival upon detachment and during metastasis .
Alterations in the expression levels of the genes of interest were confirmed by immunohistochemical staining, indicating that these molecules play an important role in the process of PTC metastasis. Laminins are heterotrimeric ECM proteins that are composed of the alpha, beta, and gamma chains, which mediate cellular signaling via interaction with cell membrane receptors. The LAMC1 gene, encoding the laminin subunit gamma 1 protein, has been reported to be involved in the progression of various malignant tumours. Previous study showed that overexpression of LAMC1 in endometrial carcinoma was related to aggressive histological types, LN metastasis, advanced International Federation of Gynecology and Obstetrics (FIGO) stage; and LAMC1 knockdown suppressed cell motile and invasive properties in endometrial cancer cells . Besides, treatment with the specific LAMC1 peptide enhanced pulmonary metastasis of B16 melanoma cells and induced the production of matrix metalloproteinase (MMP)-9 from B16 cells . Some studies have suggested that LAMC1 functions to promote metastasis and might be a novel therapeutic target in the treatment of human cancer. Consistent with these previous studies, we found that LAMC1 was expressed at a relatively higher level in samples from patients with metastasis.
THBS1 is a secreted glycoprotein involved in tumour progression via the regulation of ECM remodelling and angiogenesis. The role of THBS1 as an antiangiogenic factor is well documented; however, its effect on tumour progression and metastasis remains controversial . On the one hand, Giuseppe and colleagues found a significant reduction in THBS1 expression associated with patients presenting with thyroid carcinoma metastasis . In vitro, THBS1 could inhibit the migration of clear cell renal carcinoma cells . On the other hand, THBS1 has been reported to promote human follicular thyroid carcinoma cell invasion through the upregulation of urokinase-dependent activity  and metastasis to the lungs in a transgenic mouse model of breast cancer . Our present data indicate that THBS1 may function as a suppressor gene in PTC.
Both the gene modules MEmagenta and MEblue were inclusive of the PI3K/AKT pathway, and both of LAMC1 and THBS1 could activate this signalling pathway. Therefore, we hypothesized that this might be a key pathway involved in the process of PTC metastasis. The PI3K/AKT pathway is a key regulator of cellular processes involved in cell proliferation/the cell cycle, metabolism, apoptosis, and neovascularization . The PI3K/AKT axis could promote tumour metastasis by enhancing cell motility and angiogenesis; for instance, activated AKT could increase the bioactivities of NF-κB and VEGF, which in turn promote enhanced cell motility and angiogenesis. The PI3K/AKT pathway could also upregulate the expression of MMPs, which degrade the ECM and promote tumour metastasis and invasion .
Radiomics could provide a more accurate and robust method to predict CLN involvement in patients with PTC because the proteomic pattern is expressed in terms of image-based features. A previous study evaluated the association between the gene expression signature and CLNM in PTC . However, this molecular approach is limited by the requirement of invasive surgery or biopsy and its high cost. The imaging features in radiomic studies generally lack biological interpretations, which might hinder their clinical application. Thus, linking imaging characteristics with molecular signatures is a growing research trend that provides additional value to conventional clinical imaging with relevant molecular biological information. In this study, we identified radiomic features that were associated with gene modules, annotated the molecular and physiological effects of the relevant gene modules, and validated the associations and the results of the bioinformatic analysis, all of which lend convincing support to the proposed method. Overall, this approach highlights that imaging phenotypes allow for the noninvasive assessment of the molecular activity of PTC lesions with potential implications for treatment. Moreover, the radiogenomic map could be extended to other solid tumours and might predict the therapeutic response of existing agents through the use of gene signatures.
Several limitations were encountered in our study. First, in research on radiomics, the application of multimodal imaging is common, enlightening us to use elastography and contrast-enhanced US as other modalities to make our radiomic signature more robust and discriminative. Second, a small number of patients were enrolled in this study, and the examination of a larger cohort with available imaging, genetic and survival data is warranted. Our data come from only a single institution, and the results should be validated with a multicentre study. Finally, we focused on molecular analysis at the protein level. In the future, by incorporating other types of ‘-omic’ data, for instance, copy number variations and genetic mutations, we could provide a more complete picture of the molecular characteristics of tumours.