**6. Limitations**

Substantial radiomics studies have indicated the predictive value of radiomics in DTC; however, it is undeniable that there are also several limitations in radiomics. First, the 'black box' property of classifiers hampers the causal relationship, and the meaning of radiomics features extracted from grayscale images further hinders data interpretability. Second, radiomics is regarded as a 'population imaging' approach closely relying on different modalities and device parameters, which means variations in imaging protocols among institutions would lead to non-uniform data acquisition and thus influence generalizability. The good classification and prediction performance in a single center might not be generalized to patient cohorts from another center. Therefore, current original studies generally lack external validation. Third, although radiomics partially reflects the information at the molecular biological level, variations in tumor cells and the microenvironment as well as the retrospective nature of the studies represent limit the interpretation of the final results. Notably, based on current studies, the average diagnostic accuracy of radiomics is between 66% and 86%, even worse in the prediction of BRAF mutation, making the economic efficiency is an issue in need of attention and consideration. Furthermore, the reliability of the predictive performance and clinical application may be decreased by discussing the predictive value of radiomics itself without considering the influence of clinical information, such as tumor stages and therapy strategies. More importantly, the ethical issues regarding the use of radiomics in patient stratification and treatment response-based prognosis should also be treated with caution.
