Background
Medullary Thyroid Carcinoma (MTC) is a rare type of thyroid cancer [
1], accounting for 1%-2% of all thyroid cancers [
2]. It has a higher propensity for lateral cervical lymph node metastases, accounting for 70% [
3] of cases, as compared to other types of thyroid cancer. The American Thyroid Association (ATA) guidelines recommend surgery as the first-line therapy for definitive cure in MTC patients [
4]. The standard surgical therapy for MTC typically includes total thyroidectomy and lymphadenectomy. However, the extent of cervical lymph node dissection is still a matter of debate, particularly regarding lateral cervical lymph node dissection [
5‐
7]. The surgery decision making can affect the prognosis of MTC patients. Therefore, it is particularly important for preoperative assessment for lateral cervical lymph node metastases (LCLNM) in such patients.
Thyroid ultrasound is the first choice and a useful tool for diagnosing thyroid disease [
6]. The Thyroid Imaging Reporting and Data System (TI-RADS) has been used as a standard method for the classification of thyroid nodules [
8]. Due to its ease of use and clinical viability, TI-RADS has drawn considerable attention. The applicability of ultrasound-based TI-RADS in MTC patients has been evaluated and the relationship between ultrasound features and lymph node metastases has been assessed [
9]. However, TI-RADS results are usually affected by the experience of reviewers in most cases, and information from ultrasound imaging has not been fully explored, at present only relying on the naked eye.
Radiomics can extract quantitative features from medical images that may reflect information about underlying pathophysiology that is not visible to the human eye [
10]. In recent years, there have been numerous studies that highlight the emerging field of utilizing medical images with radiomics, combined with machine learning and deep learning to enhance the understanding and treatment of thyroid cancers, by providing personalized and detailed insights into tumor development [
11‐
13]. Biomarkers based on quantitative radiomics and deep learning features from preoperative thyroid ultrasound have demonstrated promising outcomes for predicting distant metastases in follicular thyroid carcinoma [
14], predicting thyroid malignancy [
15‐
17], and predicting lymph nodes status of patients with papillary thyroid carcinoma [
18‐
22].
Accurate prediction of LCLNM status in MTC patients can help guide surgical decisions and ensure that patients receive the most appropriate and effective surgery. To our knowledge, no studies have been published that use radiomics analysis to forecast LCLNM in MTC patients. The purpose of this study is to develop a separate biomarker which is radiomics-based for noninvasively predicting the LCLNM status of MTC using preoperative thyroid ultrasound images.
Discussion
Thyroid ultrasound is an effective tool to evaluate the lymph node status in MTC patients [
6,
8]. We investigated the association of ultrasound features from TI-RADS and LCLNM status. There were significant differences between LCLNM negative group and LCLNM positive group in margin and TI-RADS levels, and these results are consistent with the previously reported results [
9,
24]. Based on the two features, we build the regularized logistic regression model, that is TM, to predict LCLNM status. The feature importance analysis demonstrated that the margin was the more important predictor. Overall, the performance of TM as presented in Table
3 is not so good. Possible reasons may be that the TI-RADS results are commonly influenced by the experience of reviewers and more information from images have not been fully explored, as they are currently relying solely on the naked eye.
Many radiomics studies have investigated the capability for analyzing thyroid carcinoma, including prediction of distant metastases of follicular thyroid carcinoma [
14], risk stratification systems for thyroid nodules [
15], the differentiation between malignancy and benign thyroid nodules [
25], prediction of lymph nodes status [
18‐
21], prediction of BRAF mutation [
26], prediction of malignancy and pathological outcome in patients with papillary thyroid cancer [
27]. The relationship between radiomics and MTC, however, is not well understood. As far as we are aware, this study is the first to use radiomics to predict LCLNM in MTC patients. In this study, we extracted 464 radiomics features from ultrasound images and finally selected 12 features (Fig.
3) to establish the RM through a two-step feature selection approach. These features were distributed in first order statistics, GLCM, GLSZM, NGTDM and GLDM (Table
2). Some of these radiomics features, such as Small Area Low Gray Level Emphasis and Size-Zone Non-Uniformity, were similar to the previous study for prediction of distant metastasis of follicular thyroid carcinoma [
14]. The Delong test results show the RM is significantly better than the TM (
p<0.05) as shown in Table
3. Radiomics features can be taken as more powerful predictors.
Furthermore, we computed the radiomics score, referring to previous studies [
14,
15], for each patient based on the selected 12 radiomics features. We established the RTM using the radiomics score, margin and TI-RADS levels. Compared with the RM, the performance of the RTM has been further improved as shown in Table
3, Figs.
3 and
4. The radiomics score played the most important role in the RTM. The Delong test results also show the RTM is significantly better than the TM (
p<0.001). Overall, the RTM has the best performance for predicting LCLNM status in MTC patients.
In the absence of enough evidence of LCLNM, it is always questionable whether patients should undergo lymph node dissection. One of the main concerns with LCLNM is the risk of complications associated with extended lymphadenectomy, such as hypoparathyroidism and nerve palsy. Effective preoperative assessment for LCLNM status is essential. The performance of solely TI-RADS system to identify LCLNM in MTC patients is limited. Our model combining TI-RADS system and radiomics features have improved the prediction performance. It improved the ability to identify patients who require lymph node dissection while avoiding surgical complications for those who do not. This model can be used as an aid to clinical decision making when it is not clear to the clinician whether to perform lymph node dissection.
The stability and reproducibility of features are determined by ultrasound images. Ultrasound images can differ depending on the time and location they were taken, who performed the ultrasound with the probe, how much pressure was applied to the skin, patient status (e.g. body habitus, age, underlying medical conditions like skin diseases, or previous surgeries in the area), and other factors. Image normalization is a critical step that makes sure each pixel has a similar data distribution, allowing comparison with other images. Hence, the linear min-max normalization method was used to minimize this impact. On the other hand, we utilized 5-fold cross-validation to reduce the bias and maintain repeatability. Naturally, our current study lacks the validation of external data, and more data samples and involvement from more research centers will be part of our future research efforts in order to further enhance the stability of the model. Furthermore, there is often a gap between the statistical significance of radiomics features and their clinical relevance. Understanding and interpreting what these features represent in terms of underlying pathology can be challenging. This gap suggests that the practical utility of radiomics findings in clinical decision-making was only taken as a supporting tool, not as a decisive one. Additionally, the primary tumor is segmented manually, which limits the workflow's efficiency. The proposed model uses ultrasound information to help guide clinical decision-making solely from an ultrasound imaging perspective, and any other clinical information is not incorporated into the model. In the future study, a fully automated model would be developed that would include an automatic model for tumor segmentation and a radiomics-based and deep learning-based classifier comprising ultrasound and clinical information for LCLNM status prediction to further improve efficiency and accuracy.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.