Intratumoral and peritumoral radiomics based on ultrasound for the differentiation of follicular thyroid neoplasm

Introduction

Follicular thyroid carcinoma (FTC) is a common differentiated type of thyroid malignancy, which is second only to papillary thyroid cancer and accounts for about 10–15% of thyroid malignancies (1). Follicular tumors have similar cytologic, ultrastructural, and clinical features to the extent that it is difficult to determine whether they are benign or malignant (2,3). The diagnosis is mainly based on capsular infiltration, vascular invasion, extrathyroidal tumor dilation, lymph node metastasis, and systemic metastasis (3). Ultrasound (US) is still the most commonly used test and is the optimal imaging method for evaluating thyroid nodules (4). US features associated with malignancy have been shown to include a solid component, hypoechogenicity, marginal infiltration or irregularity, and the presence of microcalcifications (5). However, in other research, these features have been controversial in distinguishing follicular adenomas from adenocarcinomas (6). This may be due to the limited US features recognized by the naked eye, which restricts the diagnostic ability of US imaging. In addition, the detection rate of thyroid nodules continues to increase with the advances in diagnostic technology, especially with the improvement of US resolution, resulting in an increase in unnecessary procedures performed to establish the diagnosis (7,8). However, the detection rate of malignant tumors is low (5,9), which leads to overmedication. Overdiagnosis and overtreatment can ultimately have a serious adverse impact on the quality of life of patients and place an unnecessary financial burden on both the patient and the healthcare system (10). As a result, a nondestructive and efficient approach is urgently needed to achieve an accurate diagnosis in the early stages of the preoperative period.

Radiomics is a quantitative feature extracted from radiology images through data characterization algorithms aiming at developing prognostic prediction tools and supporting therapeutic decision-making in oncology (11). Since radiomics was proposed, a rapidly increasing number of radiomics studies have been published to improve accurate diagnosis and cancer therapeutics (12,13). Several radiomics studies have been performed on US images to classify benign and malignant thyroid nodules (14,15). To our knowledge, there are only a small number of published studies focusing on FTC using US radiomics (16,17). Besides, the published studies have mainly focused on the tumoral regions, without incorporating information from the surrounding tumor areas. Research of intratumoral images alone has not yet provided an accurate diagnosis of the disease. Since the determination of the benignity of thyroid follicular tumors correlates with the peritumoral infiltration, characterization of the peritumoral region is necessary. Therefore, the aim of our study was to investigate the efficacy of intratumoral and peritumoral radiomics in the preoperative differentiation of follicular thyroid adenomas from adenocarcinomas based on US. We present this article in accordance with the TRIPOD reporting checklist (available at https://gs.amegroups.com/article/view/10.21037/gs-24-247/rc).

Methods

Study design

Between January 2018 and October 2022, a retrospective review was performed on 195 patients pathologically diagnosed with follicular thyroid neoplasm at The Second Affiliated Hospital and Yuying Children’s Hospital of Wenzhou Medical University. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the institutional ethics committee of The Second Affiliated Hospital and Yuying Children’s Hospital of Wenzhou Medical University (No. 2023-K-43-01), and written informed consent was provided by all patients prior to operation or biopsy. The inclusion criteria were as follows: (I) patients with confirmed pathological diagnoses of FTC or thyroid follicular adenoma; (II) patients with clear B-mode US images; (III) patients with comprehensive clinical information available. The exclusion criteria were as follows: (I) patients with insufficient US image quality; (II) patients who had received preoperative therapy; (III) patients with a prior diagnosis of other malignancies or concurrent malignancies (Figure 1).

Figure 1 Study design flowchart.

Image acquisition

All US examinations were performed by US physicians specializing in superficial tissues US imaging, each having over 5 years of experience in the field. To conduct these examinations, various US machines were utilized, including the Esaote MyLab 90 X-vision (Esaote, Genoa, Italy), Toshiba Aplio 500 (Toshiba, Tokyo, Japan), and Mindray Resona 7 (Mindray, Shenzhen, China), each equipped with appropriate high-resolution probes. For subsequent analysis, maximum long-axis cross-sectional images of the targeted nodule were acquired.

All images were independently reviewed by two radiologists, each possessing more than 5 years of expertise in superficial tissue sonography, and without knowledge of any clinical information or final diagnoses. The reinterpreted US features encompassed tumor diameter, echotexture, echogenicity, boundary, margin, presence of calcification, and aspect ratio.

Lesion segmentation and feature extraction

The radiomics analysis workflow comprises critical steps: lesion segmentation, feature extraction and selection, and model construction (Figure 2). The study population was randomly divided into training and test cohorts in an 8:2 ratio. The region of interest (ROI) covering the entire lesion was manually segmented by an experienced radiologist, which was confirmed by another, with both radiologists being blinded to clinical information. The ROI was delineated through ITK-SNAP V4.0.1 software (www.itksnap.org). The peritumoral region was then generated by applying a 5 mm dilation to the delineated tumor contour, utilizing a standard morphological dilation operation implemented in a custom MATLAB program (version 2016b; MathWorks, Natick, MA, USA). As a result, 3 distinct ROI images were created for each US slice: intratumor ROI, peritumor ROI, and a combined ROI that integrated both regions. Prior to feature extraction, all US images underwent intensity normalization to standardize gray-level values. Then, 2,952 radiomics features were extracted and categorized into 3 categories: geometry, intensity, and texture. Geometry features describe the 3-dimensional (3D) shape characteristics of the tumor. Meanwhile, intensity features describe the first-order statistical distribution of voxel intensities. Texture features, derived from second- and higher-order spatial distributions of intensities, were extracted using various methodologies, such as gray-level co-occurrence matrix (GLCM), gray-level dependence matrix (GLDM), and so on. Feature extraction was carried out using Pyradiomics 2.2.0 (http://www.radiomics.io/pyradiomics.html).

Figure 2 The workflow of the radiomics model construction. ROI, region of interest; US, ultrasound; CV, cross validation; MSE, mean standard error; DCA, decision curve analysis; ROC, receiver operating characteristic; AUC, area under receiver operating characteristic curve.

Feature selection and radiomics model building

In the statistical analysis of the data, Student’s t-test was performed for features that conformed to a normal distribution, whereas the Mann-Whitney U test was used for features that did not conform to a normal distribution. We only retained features with a P value below 0.05. The correlation between features exhibiting repeatability was calculated using Spearman’s rank correlation coefficient. When the coefficient between any 2 features exceeded 0.9, 1 feature was retained. To ensure that maximal depiction of features was achieved, we implemented a greedy recursive deletion strategy for feature filtering, where we deleted the most redundant features from the current set at each iteration. Ultimately, this process led to the identification and retention of a total of 19 relevant features.

We used the least absolute shrinkage and selection operator (LASSO) method for constructing the signature within the discovery data set. The LASSO model effectively shrank all regression coefficients to zero depending on the adjustment weight λ and set the coefficients of many irrelevant features exactly to zero. To determine the optimal value of λ, a 10-fold cross-validation with a minimum criterion was performed, yielding the value of λ that produced the minimum cross-validation error. Features with non-zero coefficients were utilized for regression model matching, and combined to form radiomics signature. By linearly combining the retained features and weighting them according to their model coefficients, a radiomics score was obtained for each patient. LASSO regression modeling was performed using the Python scikit-learn package.

Following feature selection, the final features were fed into various machine learning models, including linear regression (LR), support vector machine (SVM), K-Nearest Neighbor (KNN), random forest, Extra Trees, eXtreme Gradient Boosting (XGBoost) Light Gradient Boosting Machine (LightGBM), and Multi-Layer Perceptron (MLP), respectively, for risk model construction based on intratumor ROI, peritumor ROI, and the combined ROI. The 5-fold cross verification was adopted for obtaining the final Rad Signature.

The building of clinical model and nomogram

The construction of the clinical model followed a similar process to that of the radiomics model. Features for the clinical model were selected based on baseline statistics with a P value <0.05. The machine learning model used was the same as that used for the radiomics models. Furthermore, the test cohort was set as a fixed cohort for fair comparisons, and 5-fold cross-validation was performed. An easy-to-use radiomics-clinical nomogram model was constructed integrating the radiomics signature and the important clinical predictor for potential clinical use.

Statistical analysis

To evaluate the equivalence of patient attributes in the various cohorts, independent t-tests were used for normally distributed data, and Mann-Whitney U tests were performed for non-normally distributed data. Categorical variables were assessed with chi-squared tests. The predictive ability of the model was evaluated through the construction of receiver operating characteristic (ROC) curves, calculation of the area under the ROC (AUC), and determination of the balanced sensitivity and specificity at the cut-off point that gave the maximum value of the Youden index. The AUCs between predictive models were compared using Delong’s test. All statistical analyses were conducted using SPSS software (version 21.0; IBM Corp., Armonk, NY, USA). A 2-sided P value <0.05 was considered statistically significant.

Results

Baseline characteristics

A total of 195 patients diagnosed with thyroid follicular neoplasm were enrolled for this study, comprising 153 patients diagnosed with thyroid follicular adenocarcinoma and 153 with thyroid follicular adenoma. The baseline characteristics of patients are presented in Tables 1,2. Among these clinical characteristics, only age showed a statistical difference (<0.05).

Radiomics model building and evaluation

Altogether, 2,952 features were manually extracted from both intratumoral and peritumoral regions including 611 first-order features, 27 shape features, and the remainder being texture features (Figure 3A). The comprehensive list of these features is provided in https://cdn.amegroups.cn/static/public/gs-24-247-1.xlsx. The extraction process was performed using an in-house feature analysis tool developed within the Pyradiomics framework. Figure 3B illustrates all the extracted features alongside their associated P-values.

Figure 3 Radiomics feature selection process. (A) Handcrafted features were extracted from ROIs, encompassing various categories including first-order features, shape features, and texture features (such as GLRLM, GLSZM, GLDM, GLCM). The distribution of these handcrafted features was illustrated. (B) This section displays all radiomics features along with their corresponding P values. (C,D) The optimal penalty coefficient λ for the LASSO model was generated using 10-fold cross-validation with a minimum criterion. GLRLM, gray-level run length matrix; GLSZM, gray-level size zone matrix; GLDM, gray-level dependence matrix; GLCM, gray-level co-occurrence matrix; MSE, mean standard error; ROIs, regions of interest; LASSO, least absolute shrinkage and selection operator.

We employed the LASSO method to identify nonzero coefficients for constructing the Rad-score. The coefficients along with the mean standard error (MSE) of 10-fold cross-validation is shown in Figure 3C,3D. Subsequently, 19 features were found to have nonzero coefficient values, and their specific details can be found in Figure 4.

Figure 4 Histogram depicting the rad-score distribution in terms of selected features. A total of 19 features with nonzero coefficient values were selected to develop radiomics model of intratumoral and peritumoral regions.

We obtained the optimal model by comparing rad features with various classifiers, such as LR, SVM, XGBoost, and so on. Figure 5 illustrates the AUC of each radiomics model based on the combined ROI in the test cohort. The LR model demonstrated the best performance. Therefore, we selected LR as the foundation for constructing both the radiomics and clinical signature. In the training cohort, the combined-region model yielded an AUC of 0.898 with balanced sensitivity of 0.882 and specificity of 0.856 (Table 3, Figure 6). In the corresponding test cohort, the model achieved an AUC of 0.884, with sensitivity and specificity reaching 0.750 and 0.929, respectively.

Figure 5 The AUCs of radiomics models. The picture illustrates the AUC of each radiomics model based on the combined ROI in the test cohort. AUC, area under the receiver operating characteristic curve; LR, linear regression; CI, confidence interval; SVM, support vector machine; KNN, K-nearest neighbor; GBM, gradient boosting machine; MLP, multi-layer perceptron; ROI, region of interest.

Figure 6 Diagnostic performance of different models. (A) In the training cohort, the AUCs were 0.553, 0.708, 0.842, 0.898, and 0.894 for the clinical model, intratumor-region model, peritumor-region model, combined-region model, and nomogram, respectively. (B) In the corresponding test cohort, the AUCs were 0.540, 0.710, 0.812, 0.884, and 0.884, respectively. AUC, area under the receiver operating characteristic curve; CI, confidence interval; Intra, intratumoral; Peri, peritumoral.

The intratumor-region model demonstrated an AUC of 0.708 in the training cohort, accompanied by sensitivity of 0.588 and specificity of 0.760. In the test cohort, this model had an AUC of 0.710, with the sensitivity and specificity of 0.500 and 0.857, respectively. Additionally, the peritumor-region model in the training cohort yielded an AUC of 0.842, displaying sensitivity of 0.853 and specificity of 0.760. In the test cohort, the model achieved an AUC of 0.812, accompanied by sensitivity and specificity values of 0.875 and 0.679, respectively.

Building and evaluation of clinical model and nomogram

We used the P value threshold of <0.05 to select features for establishing the clinical model in the training cohort. Age was the only characteristic that met this condition. Therefore, age was utilized in the construction of the clinical model.

The clinical model demonstrated an AUC of 0.553 in the training cohort, with sensitivity of 0.824 and specificity of 0.976. In the test cohort, it showed an AUC of 0.540, sensitivity of 0.625, and specificity of 0.964. The radiomics-clinical nomogram yielded an AUC of 0.894 in the training cohort, with balanced sensitivity and specificity of 0.853 and 0.872, respectively. In the test cohort, it demonstrated an AUC of 0.884, with sensitivity and specificity of 0.750 and 0.964, respectively (Table 3, Figures 6,7). Decision curve analysis (DCA) was conducted to assess the clinical utility of each model, with results presented in Figure 8. In comparison to scenarios without the implementation of a prediction model, the combined-region model demonstrated superior benefit across most instances (Figure 8).

Figure 7 The radiomics-clinical nomogram to predicting FTC. Sig, signature; Intra, intratumoral; Peri, peritumoral; Rad, radiomics; FTC, follicular thyroid carcinoma.

Figure 8 Decision curves of different models. DCA showing the role of 5 models in predicting FTC in the training (A) and test (B) cohorts. The y-axis indicates the net benefit, and x-axis indicates various risk thresholds. DCA, decision curve analysis; Intra, intratumoral; Peri, peritumoral; FTC, follicular thyroid carcinoma.

Statistical comparisons of AUCs among models were performed using DeLong’s test. In the test cohort, the combined-region model demonstrated statistically significant differences in AUC compared to both the intratumor-region model and the peritumor-region model (both P<0.05). The overall performance of the combined-region model was slightly higher than that of the other radiomics models. As shown in Figure 6, the nomogram exhibited best forecasting ability compared with intratumor-region model, peritumor-region model, and the clinical model, but no significant AUC differences were found between the nomogram and combined-region model.

Discussion

FTC is classified as a subtype of differentiated thyroid carcinoma. This type of cancer has a substantially worse prognosis and is about 7–23% likely to metastasize to distant organs (18). Extrathyroidal infiltration is one of the risk factors associated with distant metastases. Total thyroidectomy has represented the standard for treating FTC, however, this procedure may result in permanent hypothyroidism or potentially even hypoparathyroidism, both of which can greatly detract from the patient’s quality of life (19). Therefore, it is crucial to make an early definitive diagnosis.

Radiomics represents a precise, objective, and efficient approach to improve conventional imaging diagnostics. It involves the extraction hundreds of quantified high-dimensional characteristics from images and employs out self-training and self-learning according to pathological outcomes to aid in clinical diagnosis (20-25). Here are some published studies that apply US radiomics for identifying follicular thyroid cancer. Shin et al. (17) used US radiomics based on SVM and artificial neural network (ANN) algorithms to differentiate follicular adenoma from carcinoma, achieving AUC values of 0.605 and 0.612, respectively. Yu et al. (16) classified thyroid follicular neoplasm by combining radiomics signatures with conventional US features and clinical parameters.

To our knowledge, no previous study has applied radiomics based on signatures of the intratumoral and peritumoral region in discriminating between benign and malignant thyroid follicular neoplasm. Relying exclusively on tumor region signatures can be inadequate for achieving high accuracy, as we discovered that incorporating surrounding areas can substantially improve malignancy prediction.

In this study, we constructed and validated a predictive model consisting of intratumoral and peritumoral radiomics signature and clinical information for the preoperative prediction of thyroid follicular carcinoma by quantitative analysis of 2D thyroid US images. We compared the predictive effectiveness of each model. The combined-region model demonstrated superior predictive performance compared to the model utilizing a single region. This suggests that the potential use of US radiomics in patients with thyroid follicular neoplasm and a combined-region model would achieve more accurate predictions than an individual model.

Compared with the studies of Shin et al. (17) and Yu et al. (16), our approach demonstrated superior performance. Our study’s utilization of the linear regression model and the fact that it combined radiomics features of the inside and around the tumor could account for this finding. Besides, in the present study, the AUC of the peritumor-region model was 0.812, which showed better efficacy than intratumor-region model (AUC =0.710). This result may be explained by the characteristic of capsule or vascular infiltration around the lesion, which is related to the diagnosis of thyroid follicular carcinoma. In the study of Seo et al. (26), the model based on information of the marginal contour regions achieved strong discriminatory performance, which also supports this finding. However, their study only focused on the features of the marginal outline of the thyroid follicular tumor, while ignoring the characteristics of the intratumoral and peritumoral region of thyroid nodule images. In this study, the clinical model demonstrated a poor predictive efficacy in the differentiation of FTC (AUC =0.554). A nomogram was created on the basis of the radiomics-clinical model to help clinical decision-making. Comparison of the AUC values of the nomogram and the combined-region model further confirmed that the clinic model may not be suitable in benign-malignant differentiation. Although both the radiomics-clinical model and combined-region radiomics model demonstrated good prediction performance, the radiomics-clinical model did not outperform the combined-region radiomics model. This means that radiomics may be the most effective predictor of FTC.

Our study had some limitations. First, the sample size was limited, and data were collected from a single center. Consequently, multicenter studies with expanded sample sizes are required to further validate this model. Second, as retrospective research, there may be some selection biases that influenced the results. To address this, further prospective studies are essential to control for confounding variables. Third, US is a kind of dynamic imaging technique, and relying on a single slice may not capture all relevant information about the lesion. In future work, we will attempt to establish a radiology model based on multi-slice US images, which could provide a more comprehensive understanding of the lesions and enhance predictive accuracy.

Conclusions

This study suggested that the radiomics-clinical model consisting of an intratumoral and peritumoral radiomics signature and clinical information demonstrates satisfactory diagnostic performance for the early differentiation of FTC. The nomogram displayed good potential for clinical application.

Acknowledgments

Funding: This work was supported by the Natural Science Foundation of Zhejiang Province (No. LSY19H180008).

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://gs.amegroups.com/article/view/10.21037/gs-24-247/rc

Data Sharing Statement: Available at https://gs.amegroups.com/article/view/10.21037/gs-24-247/dss

Peer Review File: Available at https://gs.amegroups.com/article/view/10.21037/gs-24-247/prf

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://gs.amegroups.com/article/view/10.21037/gs-24-247/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the institutional ethics committee of The Second Affiliated Hospital and Yuying Children’s Hospital of Wenzhou Medical University (No. 2023-K-43-01), and written informed consent was obtained from each patient prior to operation or biopsy.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Dralle H, Machens A, Basa J, et al. Follicular cell-derived thyroid cancer. Nat Rev Dis Primers 2015;1:15077. [Crossref] [PubMed]
2. Ahn HS, Kim HS, Hong MJ. Ultrasonographic and cytologic assessments of follicular neoplasms of the thyroid: Predictive features differentiating follicular carcinoma from follicular adenoma. PLoS One 2022;17:e0271437. [Crossref] [PubMed]
3. McHenry CR, Phitayakorn R. Follicular adenoma and carcinoma of the thyroid gland. Oncologist 2011;16:585-93. [Crossref] [PubMed]
4. Alexander EK, Cibas ES. Diagnosis of thyroid nodules. Lancet Diabetes Endocrinol 2022;10:533-9. [Crossref] [PubMed]
5. Durante C, Grani G, Lamartina L, et al. The Diagnosis and Management of Thyroid Nodules: A Review. JAMA 2018;319:914-24. [Crossref] [PubMed]
6. Sillery JC, Reading CC, Charboneau JW, et al. Thyroid follicular carcinoma: sonographic features of 50 cases. AJR Am J Roentgenol 2010;194:44-54. [Crossref] [PubMed]
7. Alexander EK, Doherty GM, Barletta JA. Management of thyroid nodules. Lancet Diabetes Endocrinol 2022;10:540-8. [Crossref] [PubMed]
8. Grani G, Sponziello M, Pecce V, et al. Contemporary Thyroid Nodule Evaluation and Management. J Clin Endocrinol Metab 2020;105:2869-83. [Crossref] [PubMed]
9. De Crea C, Raffaelli M, Sessa L, et al. Actual incidence and clinical behaviour of follicular thyroid carcinoma: an institutional experience. ScientificWorldJournal 2014;2014:952095. [Crossref] [PubMed]
10. Jensen CB, Saucke MC, Francis DO, et al. From Overdiagnosis to Overtreatment of Low-Risk Thyroid Cancer: A Thematic Analysis of Attitudes and Beliefs of Endocrinologists, Surgeons, and Patients. Thyroid 2020;30:696-703. [Crossref] [PubMed]
11. Li G, Li L, Li Y, et al. An MRI radiomics approach to predict survival and tumour-infiltrating macrophages in gliomas. Brain 2022;145:1151-61. [Crossref] [PubMed]
12. Liu Z, Wang S, Dong D, et al. The Applications of Radiomics in Precision Diagnosis and Treatment of Oncology: Opportunities and Challenges. Theranostics 2019;9:1303-22. [Crossref] [PubMed]
13. Zhang J, Hao L, Qi M, et al. Radiomics nomogram for preoperative differentiation of pulmonary mucinous adenocarcinoma from tuberculoma in solitary pulmonary solid nodules. BMC Cancer 2023;23:261. [Crossref] [PubMed]
14. Zhou H, Jin Y, Dai L, et al. Differential Diagnosis of Benign and Malignant Thyroid Nodules Using Deep Learning Radiomics of Thyroid Ultrasound Images. Eur J Radiol 2020;127:108992. [Crossref] [PubMed]
15. Zhang B, Tian J, Pei S, et al. Machine Learning-Assisted System for Thyroid Nodule Diagnosis. Thyroid 2019;29:858-67. [Crossref] [PubMed]
16. Yu B, Li Y, Yu X, et al. Differentiate Thyroid Follicular Adenoma from Carcinoma with Combined Ultrasound Radiomics Features and Clinical Ultrasound Features. J Digit Imaging 2022;35:1362-72. [Crossref] [PubMed]
17. Shin I, Kim YJ, Han K, et al. Application of machine learning to ultrasound images to differentiate follicular neoplasms of the thyroid gland. Ultrasonography 2020;39:257-65. [Crossref] [PubMed]
18. Wu MH, Lee YY, Lu YL, et al. Risk Factors and Prognosis for Metastatic Follicular Thyroid Cancer. Front Endocrinol (Lausanne) 2022;13:791826. [Crossref] [PubMed]
19. Mao Y, Huang Y, Xu L, et al. Surgical Methods and Social Factors Are Associated With Long-Term Survival in Follicular Thyroid Carcinoma: Construction and Validation of a Prognostic Model Based on Machine Learning Algorithms. Front Oncol 2022;12:816427. [Crossref] [PubMed]
20. Wang Y, Liu W, Yu Y, et al. CT radiomics nomogram for the preoperative prediction of lymph node metastasis in gastric cancer. Eur Radiol 2020;30:976-86. [Crossref] [PubMed]
21. Sorrenti S, Dolcetti V, Radzina M, et al. Artificial Intelligence for Thyroid Nodule Characterization: Where Are We Standing? Cancers (Basel) 2022;14:3357. [Crossref] [PubMed]
22. Park VY, Lee E, Lee HS, et al. Combining radiomics with ultrasound-based risk stratification systems for thyroid nodules: an approach for improving performance. Eur Radiol 2021;31:2405-13. [Crossref] [PubMed]
23. Mu W, Jiang L, Zhang J, et al. Non-invasive decision support for NSCLC treatment using PET/CT radiomics. Nat Commun 2020;11:5228. [Crossref] [PubMed]
24. Liu Z, Meng X, Zhang H, et al. Predicting distant metastasis and chemotherapy benefit in locally advanced rectal cancer. Nat Commun 2020;11:4308. [Crossref] [PubMed]
25. Conti A, Duggento A, Indovina I, et al. Radiomics in breast cancer classification and prediction. Semin Cancer Biol 2021;72:238-50. [Crossref] [PubMed]
26. Seo JK, Kim YJ, Kim KG, et al. Differentiation of the Follicular Neoplasm on the Gray-Scale US by Image Selection Subsampling along with the Marginal Outline Using Convolutional Neural Network. Biomed Res Int 2017;2017:3098293. [Crossref] [PubMed]

(English Language Editor: J. Jones)