**5. Radiomics in Thyroid Cancer and Nodule Classification**

Thyroid cancer nodules are common in thyroid disease. The prevalence of thyroid nodules is approximately 67% in adults [82]. It has been reported that approximately 10% of patients with detected thyroid nodules are diagnosed with malignancy [83]. LN metastasis indicates rapid progression and poor prognosis in thyroid cancer; however, only a small portion of patients will develop metastasis [56]. Therefore, special attention needs to be paid to the differentiation of malignant and benign nodules. Table 2 presents studies focusing on the classification value of radiomics in DTC.

Machine learning or deep learning-based modalities are an important step in the processing of radiomics data. Machine learning approaches typically manually select and delimitate a set of few ROIs on the images; then, machine learning algorithms, such as support vector machine (SVM), random forest, and least absolute shrinkage and selection operator (LASSO), are applied to build a model. Prochazka et al. [84] used histogram analysis and segmentation-based fractal texture analysis algorithms combined with SVM and random forest classifiers to distinguish malignant nodules from benign nodules in ultrasound images. Their results indicated that the histogram feature was the most important parameter in classification, and both SVM (94.64%) and random forests (92.42%) achieved high accuracy. Colakoglu et al. [85] attempted to differentiate benign and malignant thyroid nodules using texture analysis and random forest model construction. After testing the reproducibility of all texture features, they finally screened seven texture features from ultrasound images, including one histogram (HistPerc 99), one HOG (HogO8b2), four GRLMs (GrlmHRLNonUni, GrlmHMGLevNonUni, GrlmNRLNonUni, and GrlmZRL-NonUni), and one GLCM (GlcmZ3AngScMom), in a random forest model. The diagnostic sensitivity, specificity, and accuracy were 85.2%, 87.9%, and 86.8%, respectively. Notably, the area under the curve (AUC) of the model was 0.92, indicating good performance. Furthermore, a recent ultrasound-image-based retrospective study recruited 2558 patients (2831 nodules), extracted radiomics features using an in-house texture analysis algorithm, and applied the LASSO method to calculate the radiomics score [86]. They used this radiomics score to determine a cutoff value that can help classify the nodules as benign or malignant. The AUCs of the radiomics score in the training and testing datasets were 0.85 and 0.83, respectively, indicating discriminative power. Furthermore, Yoon et al. [87] also applied texture analysis and the LASSO method in US images to predict malignant thyroid nodules with indeterminate cytology, demonstrating good predictive performance. Zhao et al. [88] compared the diagnostic performance and unnecessary FNAB rate for thyroid nodules of assisted visual-based and radiomic-based machine learning approaches in ultrasound images. In this study, ten machine learning classifiers, including decision tree, naïve Bayes, k nearest neighbors (KNN), logistics regression, SVM, KNN-based bagging, random forest, extremely randomized trees (XGBoost), multilayer perception, and gradient boosting tree classifiers, were verified. The results of the assisted visual-based machine learning approach indicated superior performance in AUC, sensitivity, and specificity in both the training dataset and internal validation dataset. Furthermore, a similar study design was applied to a CT-based radiomics study [89]. This study ultimately included 13 radiomics features after LASSO logistic regression. An SVM model was constructed and compared with seven other machine learning models. The study concluded that the SVM model exhibited good discrimination performance, whereas random forest had the highest stability.



203

 absolute shrinkage and

NA—not applicable.

In addition, the ability to discriminate benign from malignant lesions using a deep learning radiomics approach has also been verified by researchers. Zhou et al. [90] employed the deep learning radiomics method to differentiate benign and malignant thyroid nodules in ultrasound images. This study found that the AUC of deep learning radiomics was greater than that of other deep learning models and traditional naked-eye observations. However, the current limitation of deep learning is its black box issue, making the conclusion difficult to interpret. The abovementioned radiomic-based and deep learningbased classifications are two methods applied in the detection of malignant nodules and metastatic cervical lymph nodes. However, research comparing the diagnostic ability between these two approaches is insufficient but essential. Wang et al. [91] extracted 302 dimensional statistical features from ultrasound images and applied mutual information and linear discriminant analysis to reduce dimensionality. These researchers reported that the accuracy of radiomics for the testing data was 66.81%, which was relatively lower than that of the deep learning approach (74.69%). Although the radiomics approach may not be dominant in this study, the interpretability of deep learning is a long-standing problem that remains elusive.
