**4. Radiomics in Thyroid Cancer Prediction**

As mentioned above, radiomics aids in cancer detection, diagnosis, prediction of prognosis, evaluation of tumor status, treatment response, and local or distant metastasis [50]. Of these, the predictive value has been determined in various cancers and has been a research hotspot in recent years. Table 1 showed the predictive value of radiomics applied in DTC, Table 1 was organized according to a sequential order of prediction category, imaging method, and published time.

Metastasis is an important indicator of tumor progression [53]. Lymph node (LN) metastasis is closely related to local recurrence, distant metastasis, and thyroid stage, which further indicates the surgical plan [54,55]. Thus, the judgment of LN metastasis is important. Although a small proportion of patients report LN metastasis, those patients with suspicious abnormalities would also be suggested to undergo fine-needle aspiration biopsy (FNA) and prophylactic lymph node dissection (LND). These invasive examinations seem to be unsuitable for those people without LN metastasis. Therefore, it is important to identify a noninvasive approach to pinpoint patients with high-risk LN metastasis in clinical practice. Liu et al. [56] compared the radiomics prediction ability to estimate the LN status among B-mode ultrasound (B-US), strain elastography ultrasound (SE-US) images, and the combination of these two images. As was hypothesized, the combination group showed a better prediction ability than a single image. However, given that only 75 patients were recruited and no validation analysis was performed in this study, the results should be interpreted with caution. Furthermore, the same research team included 450 patients and divided them into training and validation datasets to verify the radiomics evaluation of US thyroid images to predict LN metastasis in PTC patients [57]. This study partly validated their previous conclusion that the features ultimately selected performed equally well

regarding the radiomics evaluation. PTC patients with or without LN metastasis showed different radiomics signatures. Jiang et al. [58] extracted radiomics features from both shear-wave elastography (SWE) images and B-mode ultrasound (BMUS) images. They calculated the Rad-score to distinguish patients with high metastasis risk. Then they built and compared the value of radiomics nomogram and clinical nomogram in predicting the LN stage. They concluded that the nomogram based on SEW radiomics signatures performed well in predicting LN status. Li et al. [59] also verified the value of ultrasound radiomics features in predicting LN metastasis. The radiomics features had a larger AUC than the ultrasound features of microcalcifications and an irregular shape.

Although CT and MRI are not exceedingly superior to ultrasound in thyroid cancer diagnosis, CT-based and MRI-based radiomics performed equally as well regarding their predictive value. The ability of CT radiomics signature to predict LN metastasis was initially reported by Lu et al [60]. This group built an SVM model and found that the radiomics signature showed a better predictive value of LN metastasis than any single radiomics signature. They concluded that the radiomics nomogram adds predictive power to LN metastasis. Hu et al. [61] initially applied multimodal MRI radiomics to predict LN metastasis in patients with PTC, and Zhang et al. [62] extracted radiomics features from T2WI and T2WI-fat-suppression (T2WI-FS) images to test and validate the predictive value of LN metastasis. These studies partly demonstrated that MRI-based radiomics can scientifically, quantitatively, and accurately predict LN metastasis in PTC patients, thereby, reducing unnecessary surgery.

LN metastasis is more likely to occur in central regions followed by lateral regions [3]. Lateral LN metastasis exhibits a higher recurrence rate and a poorer prognosis than central LN metastasis [63,64]. A recent study developed an ultrasound-based radiomics nomogram to assess its predictive value for central neck lymph node metastasis in PTC patients [65]. The prediction model showed good accuracy, sensitivity, specificity, and AUC values in both the training dataset and validation dataset. Afterward, the predictive value of ultrasound radiomics for lateral cervical LN metastasis was successively investigated in two studies. Tong et al. [66] retrospectively recruited 840 patients with PTC and extracted radiomics features from their preoperative ultrasound images. These researchers also established a radiomics-based nomogram to predict lateral LN metastasis. This radiomic nomogram presented good discrimination in both training and validation datasets and may therefore have clinical application. More interestingly, one study found a link between ultrasound radiomic features of the primary tumor and the status of lateral LN metastasis [67]. The key and interesting part of this study was that it focused on the radiomics features of thyroid primary tumors in predicting lateral LN metastasis but not the LN itself, which may facilitate the early detection of metastases.

Although the results of the abovementioned studies on the predictive value of ultrasound radiomics were largely positive in nature, the main limitation of the lack of multicenter and external validation could not be overlooked. A recent relatively robust study filled this gap. Yu et al. [3] first focused on the diagnostic value of ultrasound radiomics under a multicenter, cross-machine, multi-operator scenario. Based on B-mode ultrasound images of thyroid lesions, they established and compared four models including clinical statistical model (SM), traditional radiomics model (RM), non-transfer learning model, and transfer learning radiomics (TLR) model to predict the risk of LN metastasis in PTC patients. Of these, the TLR model showed the highest sensitivity and specificity in both the main and external cohorts. Then, a recent study that is in preprint performed an external validation based on CT radiomics indicating the good performance of this method in the prediction of LN metastasis [68]. To some extent, this study adds strength and validity to previous ultrasound-based radiomics studies.

Besides, the predictive value of radiomics was also applied in other aspects, such as the prediction of distant metastasis [69], tumor extrathyroidal extension [70,71], disease-free survival [72], and BRAF mutation [73]. The aggressiveness of tumors is classified based on various features, such as extrathyroidal extension; aggressive pathological subtypes, such

as tumors with tall cells, tumors with columnar cells, and the hobnail variant; lymph node involvement; and distant metastasis [74]. A recent study found that multiparametric MRIbased radiomics combined with a machine learning approach can accurately distinguish aggressive PTC patients from nonaggressive patients, which illustrated the role of radiomics in predicting aggressive tumors [75]. Distant metastasis of DTC is uncommon; however, FTC is more likely to have distant metastasis than PTC. It has been reported that the bone metastasis rate in FTC ranges from 7 to 28%, whereas that for PTC is only 1.4–7% [76]. Kwon et al. [69] thus evaluated the capability of ultrasound-based radiomic features to predict distant metastasis of FTC. This study is based on radiomics analysis and a machine learning approach, and multivariate analysis indicated that the radiomic signature and widely invasive histology are related to distant metastasis. Moreover, the AUC of the thyroid ultrasound radiomic signature in predicting distant metastasis was as high as 0.93, demonstrating good predictive performance. The extrathyroidal extension in patients with DTC is also an important factor to consider when determining the surgical modality. Chen et al. [70] selected five CT-based radiomics features that were closely related to the extrathyroidal extension of PTC patients. A CT-based radiomics nomogram was built and showed good predictive value in extrathyroidal extension. This excellent predictive performance for tumor extrathyroidal extension was also verified in an MRIbased radiomics preprint [71]. Regarding "disease-free" cancers, DTC has an overall good disease-free survival after treatment and long-term outcomes [77]. Despite being called a "happy cancer", tumor progression contributes to the 1.4–5.2% mortality rate of thyroid cancer [78,79]. A retrospective study included 768 PTC patients, extracted radiomics features from ultrasound images, and constructed a radiomics signature based on LASSO regression. Finally, a Rad-score was calculated to stratify the patients into high- and lowrisk DFS [72]. Furthermore, based on recent progress in molecular genetics, gene-specific information has provided insights into the biology of the tumor, prediction of prognosis, and potential therapeutic targets [80]. The B-Raf proto-oncogene serine/threonine kinase (BRAF) mutation is involved in the pathogenesis of PTC and is related to tumor progression, recurrence, and mortality [73]. In addition, shedding light on the mutational status of thyroid cancer could help clinicians evaluate the tumor response to new drugs, such as tyrosine kinase inhibitors. Thus, if we can predict genes mutated in thyroid cancer through convenient and feasible approaches, this information would contribute to improving tumor diagnosis, judging the prognosis, and personalizing the treatment. To date, two studies have applied radiomics to estimate BRAF mutations in PTC patients [73,81]. These two studies offered a consistent outcome that ultrasound radiomics has a limited value in predicting BRAF nutation. This result indicated that the relationship between ultrasound radiomics and gene mutation may not be as good as expected.

*Cancers* **2021**, *13*, 2436


0.93(IT2)

*Cancers* **2021**, *13*, 2436


200

*Cancers* **2021**, *13*, 2436


SWE—shear-wave elastography; CT—computer tomography; SVM—support vector machine; RF—random forest; LASSO—least absolute shrinkage and selection operator; CV—cross-validation; LOOCV—leaveone-out CV; EV—external validation; TLR—transfer learning radiomics; SM—statistical model; RM—traditional radiomics model; NTLR—non-transfer learning radiomics; IT—independent set; DM—distance metastasis; MRI—magnetic resonance imaging; ETE—extrathyroidal extension; DFS—disease-free survival; LSVM—linear support vector machine; LR—CV-logistic regression classifier with cross-validation; PAC—passive aggressive classifier; LSVC—linear support vector classification; mRMR—minimum redundancy maximum relevance; NA—not applicable.
