Investigating the association between radiomic and genomic data is a complex multi-step process that could represent a new approach for the definition of diagnostic, predictive, or prognostic cancer models [31, 32]. One of the fundamental goals of this interdisciplinary approach is to identify specific radiological patterns that correlate with the genomic characteristics of tumors [50].

Using imaging techniques, such as US, MRI, CT, along with sophisticated computational algorithms, clinicians can extract quantitative data from images [30, 51, 52]. The extracted radiomic features serve as a bridge to connect with genomic data, obtained through high-throughput biological techniques (e.g., NGS, Sanger sequencing).

In the context of TC, PTC, the most prevalent variant, generally exhibits a favorable prognosis [4, 5]. However, there are high-risk cases that may undergo recurrence and metastasis, with lymph nodes, both central and lateral, serving as primary sites. Furthermore, PTCs are marked by intratumor heterogeneity, which has a detrimental impact on the prognosis, particularly in advanced tumor stages [53]. Therefore, understanding crucial genetic events in tumorigenesis and correlating them with imaging data, could improve diagnosis and prognosis and thus contribute to a better management of patients.

Based on these assumptions, this systematic review was aimed to highlight the potential applications of the correlation between genomics and radiomics in TC. Through a comprehensive analysis of literature, a total of 7 studies were considered eligible and were grouped into two main categories based on clinical outcomes: predicting metastasis or predicting mutation status.

Thus, in this systematic review the application of radiomics in the prediction of lymph node metastasis was investigated on the basis of the three reports aforementioned [40, 45, 46]. Dong's paper [46] explores a CT radiomics-based nomogram for predicting LNLN metastasis in PTC. There are many strengths of the nomogram obtained from CT such as: (i) the selection of lymph nodes as segmentation objects of research; (ii) the robustness; (iii) stability and performance comparable to senior radiologists. Nevertheless, as highlighted by the authors, there are several limitations: (i) the sample size is too low, (ii) the manual delineation of the ROIs, (iii) the correlation between the CT images with pathological reports can be subjective and uncertain; (iv) experimental validation is needed to support the observation that radiomic features reflect changes in the tumor microenvironment; (v) sequencing was performed on a few tissue biopsies.

The paper of Tong [40] aims to predict CLNM in PTC patients using preoperative thyroid US radiomics and gene modules by WGCNA, highlighting the potential of radiomics compared to the standard imaging assessment, and the correlation with specific gene pathways. Nevertheless, this report has also limitations, such as single-center imaging data and non-compliance to the Imaging Biomarker Standardization Initiative.

The heterogeneity of PTC, characterized by various gene mutations and fusions, necessitates advanced diagnostic tools for better treatment planning. NGS has become instrumental in identifying these genetic changes, which are crucial for predicting aggressive cancer phenotypes.

In the study by Zhang and colleagues [45], NGS was employed to compare genetic mutations (e.g., BRAFV600E) in PTC tissues and benign nodules. The authors found that combining gene alterations with ultrasound-related radiomics features improved predictions of lymph node metastasis. Despite these promising findings, the study has several limitations, including a sample size that may not be large enough to generalize the results to all populations. Furthermore, although the study provides a predictive model, it requires long-term validation to confirm its accuracy and reliability in clinical settings.

Thus, overall, these articles highlight the promising role of associating radiomics with genomics data as a non-invasive tool in enhancing TC diagnosis and prognosis. However, addressing these limitations in future research will be crucial to enhance the robustness and clinical applicability of the models.

The other papers included in this systematic review examined the combination of radiomic and genomic characteristics in predicting mutation status. Currently, the mutation status is determined through FNA and gene sequencing [54, 55]. Despite its widespread use, FNA can give false-negative [56]. Hence, a non-invasive preoperative screening tool and accurate knowledge of mutation status could be necessary for diagnosis of PTC, providing clinicians novel strategies to optimize therapeutic management.

In Zheng et al.’s work [41], the texture model, utilizing MRI features, demonstrates promising predictive value for BRAFV600E mutation in PTC. Despite that, this study acknowledges limitations, including relatively small sample size, potentially leading to overfitting. Furthermore, the study excluded thyroid lesions smaller than 5 mm and focused only on the largest lesions in patients with multiple PTC.

Also, in the study of Wang and colleagues, US radiomic models based on elastic features and the combination of both the elastic and grayscale features revealed to be a potential tool to assess BRAFV600E mutation status in PTC [42]. Several limitations have been reported in this study, such as: (i) the small sample size, which may lead to selection bias; (ii) radiomic features were extracted from images generated by only two ultrasound devices; (iii) the lack of genetic analysis of the BRAF gene in healthy subjects and, finally, (iv) the absence of external validation data.

On the contrary, in Yoon’s study [43], radiomic features extracted from US did not prove to be reliable for the preoperative diagnosis of the BRAFV600E mutation status in patients with PTC, regardless of tumor size, but only in a subgroup analysis of tumors sized less than 20 mm. Furthermore, the main limitation of this study was related to the inherent observer variability of US, and also that the ROI segmentation was performed by one radiologist, which may have influenced the results. Another limitation may be associated with the enrolled population, which may not accurately reflect the mutation features of the TC population.

In the same context, the performance of a deep learning radiomics nomogram derived from the combination of US radiomics and deep transfer learning (DTL) in preoperative RET/PTC rearrangement detection was investigated [44]. The results showed a valuable predictive ability of the radiomics nomogram and demonstrated that DTL can potentially be integrated into diagnostic procedures to assist clinicians in their medical decisions. However, some limitations have emerged from this work and further studies are needed to validate the predictive ability of the model in a larger sample size and to extend the field to the specific RET rearrangement subtypes.

The predominant molecular alterations observed in PTC are associated with the activation of the MAPK signaling pathway, which includes point mutations in BRAF and RAS genes, and fusions involving the tyrosine kinases RET and NTRK1. Additionally, mutations in PTEN, PIK3CA, and AKT1 genes have been identified with less frequencies [10, 17]. Thus, despite a thorough understanding of the molecular profiles associated with PTCs, the literature lacks studies that utilize radiomic signatures to predict other mutations, such as those in RAS or other key genes, in addition to BRAF and RET. This gap highlights a critical area for future investigation.

Overall, from the analysis of TC radiogenomics literature, a substantial scarcity of work emerges which, although well evaluated by the quality control used in the review, does not allow a homogeneous evaluation and, above all, a comparison of the results.

We could not compare our systematic review with others on this topic because to the best of our knowledge, this is the first review that investigates the potential association of radiomic features with genomic signatures in TC.

On the other hand, other tumor districts, such as colon [57], breast [58,59,60,61,62,63,64,65,66], prostate [67], lung [68, 69], bladder [70], gliomas [71, 72] of comparable incidence, have favored the production of a large literature. In these cancers quantitative radiomic features associated to the genomic signature were able to reveal tumor heterogeneity, to distinguish molecular subtypes and to unveil novel potentials therapeutic targets. This difference can be possibly, mainly attributable to the imaging methods used in diagnostic practice and the related data made available to the scientific community. In fact, TC is characterized by the prevalent use of US which, due to its intrinsic limitations, does not allow optimal radiomic analysis.

Future research could take advantage of the data deposited in The Cancer Imaging Archive (TCIA). The TCIA, in connection with The Cancer Genome Atlas (TCGA), is a useful tool for radiogenomics providing datasets that combine imaging data with genomic information, helping clinicians in the development of more accurate diagnostic and therapeutic strategies.

The interplay between radiomic and genomic features plays an important role in the advancement of diagnostic, predictive or prognostic tumor models, representing a promising avenue in cancer research [73]. For this reason, the development of both novel nomograms and genomic methods and their association could be useful for the advancement of personalized medicine.

A major limitation of our work is the low number of studies (n = 7) that meet the search criteria on association of radiomics and genomics. Indeed, most of the studies included in the search focus on radiomics underlining the potential of radiomic signatures as non-invasive tool, providing clinicians additional information for preoperative decision-making and prognostic assessment. Notably, the ability to predict lymph node metastasis in PTC using radiomic features from various imaging modalities, such as CT and US, indicates a promising direction for improving patient management strategies.

Another important limitation of this systematic review is certainly the methodological heterogeneity observed among the different studies discussed in the field of radiomics. Indeed, different imaging modalities, such as US, CT and MRI were utilized and this methodological heterogeneity is also evident from genomic methods, where the use of sequencing techniques (i.e., NGS, Sanger sequencing) and PCR introduces an additional layer of variability.

While US is the most used imaging technique as a stand-alone approach to TC, representing the gold standard with FNA for the diagnosis, it represents the less reliable, sonographer-dependent approach compared to other cited modalities. However, considering the widespread and the cost of this imaging technique, more efforts have to be spent to increase its standardization and reproducibility.

Furthermore, the use of different software and methods for generating features complicates the translation of the concept of radiogenomics into clinical practice. In clinical settings, reproducibility and reliability of measures and procedures are crucial. Consequently, the interpretation of results may be influenced by this methodological heterogeneity, highlighting the need for a methodological standardization process.

These characteristics gain more importance if it is considered that the small sample size represents the main common drawback of all examined studies. Finally, although intrinsically heterogeneous, the PTC is the more common TC but not the unique subtype, limiting the generalizability of these findings.

Additionally, one limitation of this systematic review regards the different design of the studies included. Indeed, out of the 7 studies, 1 was prospective, 5 retrospective and 1 was prospective and retrospective multicenter study.

Another limitation of our review could be the bias towards positive findings. Indeed, 6 out of 7 studies included in our analysis state that the association of radiomic and genomic studies is beneficial. Only one study suggests that the association of US-radiomics data and genetic mutations has limited predicting value [43]. The bias towards positive findings could be due to the low number of studies that are available on this topic.

In order to overcome, or mitigate, all these limitations, the use of well-established tools in the bioinformatics field, such as multi-assay experiments, which have recently been demonstrated to be suitable by integrating the radiomics domain [74, 75] should certainly be favored and promoted.

In conclusion, while the synergy between radiomic features and genomic signatures holds promise as a non-invasive tool for enhancing TC diagnosis and management, further researches, with larger cohorts and standardized methodologies, are essential to establish its clinical utility.