Our study combined the clinical factors, radiomics features and frozen section to build nomogram as a new and invasive approach for predicting LNM preoperatively in PTC. The results showed that the nomogram had favorable discrimination in both training group and validation group and it could facilitate the individualized prediction of LNM for the patients with PTC.
Previous studies have found that LNM was associated with local recurrence, distant metastasis and poor survival [6, 7, 33]. The status of LN could impact the treatment decisions. To some extent, prophylactic LN dissection would decrease the risk of poor outcomes, while increase the risk of hypoparathyroidism and nerve injury. Therefore, to reduce the number of preventative LN dissection and help clinicians build personalized surgical strategies for patients, the status of LN should be accurately diagnosed. However, the accuracy of US-reported LN status was not satisfied in clinical practice, as it was difficult to detect LNM including small and nontypical LNs. Previous studies have evaluated the ultrasonographic signs of thyroid nodules such as tumor size, echogenicity, calcification, “wider than tall” and extrathyroidal extension were associated with LNM [33–35]. FNA was regarded as the most direct way to increase the diagnostic performance before operation [36]. Although the mentioned approach above of examining LNM were encouraging, the diagnostic accuracy was largely depended on the experience of the radiologists.
A noninvasive and effective approach should be evaluated for the accurate diagnosis of LNM. Radiomics analysis was reported to have the potential for achieving personalized medicine across different cancer types [13, 37]. However, its potential was less investigated in predicting LNM for the patients with PTC. Several studies have focused on the radiomics features associated with the status of LN to provide better predictive model in clinical practice [14–23]. Li et al.[15] found that radiomics analysis based on the 150 thyroid nodules had the AUC of 0.759, with a sensitivity of 0.90 and specificity of 0.860 in the training set, while the AUC of 0.803, with a sensitivity of 0.727 and specificity of 0.800 in the validation set. Another study extracting 50 radiomics features based on 450 patients with PTC to predict LN status showed the AUC of 0.782 and accuracy of 0.712 in the validation cohort, the AUC of 0.727 and accuracy of 0.710 in the independent testing cohort [16]. The results of our study were similar to other studies that predicting the LNM in patients with PTC. Our study found that the sensitivity, specificity and AUC of radiomics model were 63.6%, 77.5%, 0.702 (0.617,0.788) in the training group, 76.5%,53.6%, 0.616 (0.617,0.788) in the validation group. Although the performance of discrimination was slightly lower in the validation group, US radiomics features were strongly associated with LNM for patients with PTC. Considering the performance of radiomics model, we further selected clinical characteristics and frozen section that were strong predictors to build nomogram for predicting LNM.
Intraoperative FS was advocated as a useful method for surgical planning, as it could provide more precise diagnosis. Previous studies illustrated that FS combined clinical and radiological features showed good performance and diagnostic accuracy [38–40]. Wu et al.[39] investigated a CT-based radiomics model and combined it with FS and clinical data to distinguish invasive adenocarcinomas (IA) from preinvasive lesions/minimally invasive adenocarcinomas (PM). The results showed that the model yielded AUC of 0.96, 0.97 and 0.96 in the training, testing and external validation dataset. Sun et al.[38] evaluated the diagnostic performance of radiomics model and FS for the pathological classification of peripheral lung adenocarcinoma manifesting as ground-glass nodules (GGNs) in CT. The results found that the concordance rate between FS and final pathology when FS had the same outcome with radiomics classifier was significantly higher than when FS had different outcome from radiomics classifier. The study indicated that the diagnosis when the outcome of FS was different from radiomics methods should be considered seriously. Considering the FS of intraoperative LN remained false negative, FS of thyroid cancer have the potential ability to predict the status of LN. Previous studies have found PB was significantly associated with local recurrence and poor survival[27, 28, 41]. A study based on 258 PTC patients reported that the presence of PB was significantly associated with LNM [42]. The results were similar to our study, as PB was strongly related to LNM in the univariate and multivariate logistic analysis.
To best of our knowledge, previous studies have not incorporated the clinical characteristics, FS and radiomics features to build model for the discrimination of LNM in patients with PTC. Our study had developed the clinical model, radiomics model and nomogram to predict the status of LN. The results showed that the nomogram had good predictive value with AUC of 0.822 in the training group and 0.803 in the validation group. It had minor non-significant improvements in AUC compared to clinical model and significant improvements compared to radiomic model, however, the sensitivity of nomogram was a little higher than clinical model and radiomics model. Compared to the radiomics model, the diagnostic performance of clinical model was not significantly different in the training and validation group. Nomogram was the firstly recommended for the prediction of LNM in patients with PTC, as it could provide higher diagnostic efficiency. Incorporating the frozen section and radiomics analysis provided new insights in predicting the status of LN for patients with PTC. Our study recommended that the nomogram could be used to predict LNM before surgery in hospitals with the conditions for conducting the research of radiomics. It would hold promise for personalized treatment and facilitate the decision making in clinical practice.
Our study had several limitations. First, our study was retrospective and the selection bias could not be avoided. Second, the sample size was relatively small based on single-center study. Third, the radiomics features were influenced by the quality of US images that associated with different US machines and the experience of radiologists. Fourth, we did not divide the cervical LN into central and lateral LN. Due to the sample size, most PTC patients had central LNM with/without lateral LNM. Further study based on multi-center large sample size are needed to verify the reliability of our model.