The incidence of PTC is rising progressively, with an increasing number of patients diagnosed worldwide. Surgical resection of the thyroid tumor region is recognized as the preferred treatment method. However, whether pCCND is needed simultaneously for all patients with PTC who are clinically node negative (cN0) remains debatable. According to the 2015 ATA guidelines, performing pCCND for all patients with PTC is not advisable, especially for those with small primary (T1 or T2) and noninvasive tumors(20). However, in China, the latest guidelines for differentiated TC recommend that, while preserving the parathyroid gland and the recurrent laryngeal nerve, at least ipsilateral pCCND should be performed༈21༉. In this study, 64 of the 229 patients were confirmed to have CLNM by postoperative pathology, accounting for approximately 28% of the overall sample, a percentage that is consistent with the literature༈22༉. Therefore, even when the PTC surgery is conducted by highly specialized surgeons and experts, clinicians should carefully access the relative risks and benefits of pCCND for every PTC patient. To accurately guide patients in selecting the appropriate surgical method, thus avoiding overtreatment and reducing individual and medico-economic burden, there is a pressing need to develop a practical predictive model to enhance the preoperative predictive accuracy of CLNM༈20༉.
US images of most patients were recognized as the best choice to diagnose PTC, although they may not reveal any abnormal findings about CLNM preoperatively. The US examination is not reliable for visualizing deep anatomic structures, especially for objects acoustically shadowed by air and bone(23). Furthermore, owing to the subjective nature of the US examination, which is based heavily on the sonographer’s experience, it can result in variability between observers and may affect the accuracy of the CLNM diagnosis༈24༉. Radiomics is a novel and noninvasive method, which extracts and analyzes medical image characteristics according to tumor heterogeneity to establish a predictive model to improve diagnostic and predictive capacity༈25༉. USR techniques have rapidly advanced and have found application for the differential diagnosis of tumors, including the prediction of microvascular invasion in hepatocellular carcinoma༈26༉, for the discrimination of high-risk endometrial cancer༈27༉, and for the evaluation of breast cancer response to chemotherapy༈28༉. These radiomic features hold promise as noninvasive biomarkers for predicting CLNM in patients with PTC. The advancement of computing power has fostered the development and growth of artificial intelligence applications, which can provide significant assistance to humans in tackling complex decision-making tasks༈29༉. However, various algorithms utilize distinct principles for constructing prediction models. Therefore, finding a compatible algorithm to improve the accuracy and diagnostic capacity of radiomic models is becoming the focus of attention༈19, 30༉.
In our study, artificial intelligence ML algorithms, five models of which were constructed to assess preoperative prediction of CLNM in patients with PTC based on US radiometric characteristics filtered by LASSO, were compared. Ultimately, we selected 10 radiomic signatures using LASSO regression. The parameters included are shown in Table 3. All the chosen radiomic features were wavelet-based, capable of uncovering hidden information within medical images across multiple scales(31).
We compared the MLMs established by the RF, DT, and AdaBoost algorithms, although we generated a marked overfitting problem. RF, DT, and AdaBoost are tree-based non-linear algorithms that are efficient and accurate methods for variable selection and classification(32); however, these classifiers result in robust noises and outliers, causing an overfitting༈33, 34༉. This is in accordance with the research findings reported by Yin et al.༈17༉. The performance of SVM and LR models showed better stability than RF, DT, and AdaBoost in the training and validation sets. In the training sets, the AUC of the SVM model (0.890) slightly outperformed that of the LR model (0.831). However, in the validation sets, the RF model exhibited superior performance compared to the SVM, achieving the highest values for both accuracy (0.735) and F1-score (0.795) among the five classifiers. In general, the results of this study reveal that the AUC values demonstrated the LR-based ML model's ability to distinguish between CLNM-positive and CLNM-negative patients with PTC in both the training (AUC: 0.831) and validation (AUC: 0.722) sets, signifying a satisfactory model performance. Therefore, LR was established as the CLNM prediction model for PTC in this study. Thereafter, a nomogram, utilizing the Rad_score, was created to visually represent the LR model.
The LR model in this study demonstrated a higher predictive accuracy than some models established based on traditional clinical features in previous studies(30). Agyekum et al.༈6༉ developed a CLNM prediction model incorporating clinical risk and USR, employing the LR algorithm. The AUC (0.710) in the validation set for the CLNM diagnosis model was slightly lower than that observed in the present study. In the study by Li et al.༈35༉, the authors created a computer model based on deep learning for CLNM diagnosis in patients with PTC, and the AUC of their validation sets (0.794) was higher than that observed in our study (0.722). This divergence could be attributed to the superior capacity of deep learning algorithms to extract high-level features from datasets compared to traditional ML algorithms. Zhou et al.༈36༉ developed a USR nomogram for preoperatively predicting CLNM in 609 patients; the AUCs in the training and validation sets were 0.816 and 0.858, respectively. The high value of AUC could be related to the larger sample size of patients included in the study and the integration of USR features with clinical features. Furthermore, the LR model showed excellent and consistent performance in data processing and ML prediction of various diseases, such as prediction of pulmonary nodules༈37༉ and prediction of breast cancer invasiveness༈38༉.
This study had several strengths. First, the strict enrollment criteria and patient inclusion avoided interference due to bilateral PTC lesions leading to bilateral CLNM. Next, to mitigate selection bias, this study employed a completely randomized grouping design and conducted consistency assessments among observers and intra-observers. Lastly, the USR model established based on the LR algorithm was a simple diagnostic and prognostic tool, assisting PTC patients without CLNM in avoiding unnecessary surgery.
Nonetheless, this study had few limitations. This was a retrospective study; therefore, the results may be influenced by a case-selection bias. Furthermore, the sample size was limited, and all cases were collected from a single hospital. Additionally, while we demonstrated the potential feasibility of applying the ML model and incorporated US-based radiomic data to predict CLNM in patients with PTC, our study was further limited by the absence of external validation. Therefore, future prospective studies with a larger number of samples and centers are required to verify this model.