Dual-channel ultrasonic images empowered deep learning: significantly improving prediction of occult central lymph node metastases in solitary papillary thyroid microcarcinoma

Papillary thyroid carcinoma is the most prevalent malignant tumor in the endocrine system.¹ Over the past decade, its incidence has been rising significantly.² Papillary thyroid microcarcinoma (PTMC) is a special subtype of papillary thyroid carcinoma. It has a maximum diameter of no more than 1 centimeter and makes up 60%–90% of newly diagnosed papillary thyroid carcinoma (PTC) cases. While the overall 10-year survival rate for PTMC patients is relatively high, about 13.5%–35.6% of them experience cervical lymph node metastasis.^3–5 Among those with cervical lymph node metastasis, roughly 85% present with central lymph node metastasis (CLNM). This pathological manifestation significantly elevates the risk of local recurrence after surgery⁶, thereby influencing clinical treatment decisions and prognostic management.

The role of prophylactic central lymph node dissection in clinically node-negative (cN0) PTMC remains contentious globally, with persistent divergences in risk-benefit evaluations among surgical societies.^7–9 Recent clinical developments have introduced non-surgical alternatives (active surveillance and thermal ablation) for managing low-risk thyroid cancers. While these approaches show favorable prognoses with preserved organ function, 2.1% to 2.6% of patients ultimately develop CLNM, highlighting the need for refined risk assessment.¹⁰ This clinical progression frequently necessitates additional surgical interventions and potentially adjuvant radioactive iodine therapy, which may significantly compromise both physical and psychological well-being in affected individuals.

Precision risk stratification for occult CLNM represents a critical determinant in personalizing PTMC management. Current prediction algorithms primarily combine ultrasound imaging features with clinical indicators; however, they still face significant challenges in predictive accuracy due to interobserver variability in image interpretation and biological heterogeneity.¹¹ The exponential growth of artificial intelligence in medical imaging has catalyzed the application of DL architectures for CLNM prediction in PTMC.¹² However, the current body of evidence highlights a critical gap in the availability of rigorously validated artificial intelligence based predictive frameworks specifically tailored to detect occult CLNM.

In this study, we have developed an innovative deep neural network model. This model is constructed using dual-channel fused ultrasound images, incorporating both transverse and longitudinal sections, and is further integrated with clinical characteristic data. The primary objective of this model is to provide accurate preoperative predictions of occult CLNM in patients with PTMC. By achieving this, the model aims to support decision-making processes in selecting individualized treatment strategies for PTMC patients.

A total of 496 patients were initially selected for this study. However, 34 patients were excluded due to incomplete ultrasound image data, and 1 patient was excluded due to confirmed CLNM before surgery. According to the research design and criteria, these 35 patients were excluded (the specific exclusion process is shown in Figure 2). Finally, 461 patients were included in the study. The average age of the patients was 44.5 ± 11.6 years, with females comprising 74.8% (345/461) and males comprising 25.2% (116/461). In total, 922 ultrasound images were used in the study. Among these, 758 images from 379 patients were allocated to the training and validation sets, which were used for model development and internal validation. Additionally, 164 images from 82 patients were used for external testing to assess the model’s generalization ability. Statistical analysis revealed no significant differences in clinical indicators between the training set, validation set, and external test set (see Table 1 for details), with P values ranging from 0.016 to 0.900.

FIGURE 2.

Flowchart illustrating the patient enrollment process.

PTC = papillary thyroid carcinoma; PTMC = papillary thyroid microcarcinoma; CLNM = cervical lymph node metastasis

Performance of the Dual-channel Model

In both the training and validation sets, the AUC values of Model 3: DL_F were 0.765 (95% confidence interval [CI]: 0.703 - 0.826) and 0.723 (95% CI: 0.617 - 0.828), respectively. These values were significantly higher than those of Model 1: DL_L (0.673 [95% CI: 0.603 - 0.742] and 0.671 [95% CI: 0.563 - 0.779], P = 0.049, 0.019) and Model 2: DL_C (0.618 [95% CI: 0.542 - 0.693] and 0.542 [95% CI: 0.426 - 0.658], P = 0.002, 0.042). In the external test set, the AUC value of Model 3 was 0.726 (95% CI: 0.574–0.878), which was also higher than the AUC values of Model 1 (0.547 [95% CI: 0.400-0.691]) and Model 2 (0.510 [95% CI: 0.359-0.662], P = 0.043, 0.049). Additionally, gradient-weighted class activation mapping (Grad-CAM) was used to perform an interpretability analysis of Model 3 (Figure 3).

FIGURE 3.

Gradient-weighted Class Activation Mapping images predicted by the dual-channel deep learning model. The left image represents the minimum bounding rectangle containing the largest section of the tumor on the longitudinal section. The middle image represents the minimum bounding rectangle containing the largest section of the tumor on the transverse section. The right image represents the Gradient-weighted Class Activation Mapping image.

Establishment and performance evaluation results of the combined model and model visualization

For the fused dual-channel images, this study extracted 2048 DL features, from which 64 compressed features were obtained and integrated with all clinical indicators. After multiple rounds of feature screening, 7 DL features and 4 clinical features (age, gender, nodule size, aspect ratio) were retained. The LightGBM classifier demonstrated higher AUC values in both the training set (0.900 [95% CI: 0.861–0.940]) and the validation set (0.862 [95% CI: 0.778–0.947]) compared to other classifiers and was selected as the final model for the combined model (Model 4). The AUC value of Model 4 in the external test set was 0.873 (95% CI: 0.774–0.972) (Figure 4).

FIGURE 4.

The AUC values of the combined model. Compared with the clinical model and the dual-channel model, the combined model has higher AUC values on the training set (A), the validation set (B), and the external test set (C).

AUC = area under the receiver operating characteristic curve; CI = confidence interval; LightGBM = Light Gradient Boosting Machin; LR = logistic regression

In both the training and validation sets, the AUC of Model 4 was significantly higher than that of Model 3 and Model 5 (training set: all P < 0.001; validation set: Model 4 vs. Model 3, P = 0.042; Model 4 vs. Model 5, P = 0.012). In the external test set, the AUC of Model 4 was higher than that of Model 5 (P = 0.013). However, the difference between Model 4 and Model 3 was not statistically significant (P = 0.122). Calibration curves indicated that the prediction performances of Model 3, Model 4, and Model 5 were all satisfactory (all P values > 0.05) (Figure 5). The SHAP value graph illustrated the contributions of DL features and clinical features to Model 4 (Figure 6).

FIGURE 5.

Calibration curves of different models. (A) Calibration curves of the combined model, dual-channel model and the clinical model in training dataset; (B) Calibration curves in the validation dataset; (C) Calibration curves in the external test set.

FIGURE 6.

Contributions of individual features in the combined model. The summary plot displays the distribution of SHAP values for each feature across all predictions in the training cohort, providing a comprehensive overview of feature importance and their respective impacts.

SHAP = SHapley Additive exPlanations

The objective of this study was to develop an AI model for preoperative prediction of occult CLNM in PTMC. A retrospective analysis of 922 ultrasound images from 461 patients with PTMC was conducted, employing DL and various machine learning techniques to build the AI prediction model. Key findings from this research demonstrated that the dual-channel model outperformed the individual longitudinal-axis and transverseaxis DL models in predictive performance. Furthermore, its accuracy was further enhanced when integrated with clinical features. These findings hold significant clinical implications, as the model can aid physicians in developing more personalized treatment and management strategies for patients with low-risk thyroid cancer, thereby improving patient prognosis and reducing unnecessary medical resource utilization.

In this study, we observed that the AUC value of Model 3 was significantly higher than those of Models 1 and 2 across all datasets. Specifically, when compared to individual DL models based on transverse and longitudinal sections, the dual-channel model demonstrated superior performance in predicting occult CLNM. The dualchannel model integrates both transverse and longitudinal ultrasound image data of the tumor. Its unique dual-channel architecture allows for efficient extraction and fusion of features across multiple dimensions, enabling comprehensive capture of imaging characteristics associated with occult CLNM. This capability is central to its enhanced predictive performance. In contrast to previous studies, which predominantly focused on ultrasound images from a single section, the dual-channel model used in this study introduces a novel perspective and approach to the field.^12,15 These findings not only expand the technical methods for predicting occult CLNM in patients with PTMC but also offer new insights and directions for future research. Clinically, the high AUC value of this model suggests that it can more accurately predict CLNM prior to surgery, facilitating more precise preoperative assessments and enabling the formulation of more tailored surgical strategies for patients.

In constructing the combined model for this study, we applied multiple feature dimensionality reduction methods, ultimately retaining four clinical features: age, gender, nodule size, and aspect ratio. SHAP value analysis revealed that younger patients, male patients, and those with larger nodules were more likely to experience occult CLNMs, which aligns with findings from numerous previous studies.^16,17 In PTMC, younger patients and males are at a higher risk for occult CLNM, with tumor size widely recognized as a key factor influencing this process. The underlying mechanism is attributed to the enhanced proliferative and invasive phenotypes of tumor cells in younger patients, thereby enabling more efficient invasion of local tissues and penetration of the lymphatic vasculature, ultimately culminating in lymph node metastasis. Additionally, male hormone levels and immune status may play a role in tumor development and metastasis, making men more prone to metastasis. Larger nodules are also more likely to invade surrounding lymphatic and blood vessels, providing greater opportunities for metastasis. This study not only corroborates these existing conclusions but also, through rigorous feature dimensionality reduction and SHAP value analysis, provides a more robust quantitative foundation for understanding the significance of these factors in predicting metastasis. However, analysis of the SHAP value graph revealed that the relationship between the aspect ratio and SHAP value could not be determined due to considerable overlap between nodules with higher and lower aspect ratios among the patients with occult central lymph node metastasis.

In this study, we extracted DL features from dual-channel fused images and integrated them with clinical features to develop a combined model. This model demonstrated exceptional performance across different datasets, with its prediction results in both the training and validation sets significantly outperforming those of the single dualchannel model and the clinical model. In previous studies, Wang et al. employed a single DL model to predict occult CLNM in PTMC. Although this model successfully extracted image features, it lacked the incorporation of clinical factors, which hindered its ability to comprehensively and accurately predict metastasis risk, especially when considering diverse individual patient characteristics.¹² In contrast, primarily relied on clinical features to construct a prediction model. While this approach accounted for overall patient condition, it was limited by its inability to extract latent information from imaging data.¹⁸ Similar to the clinical model in our study, this approach overlooked crucial imaging-level information when faced with complex and variable imaging manifestations, leading to suboptimal prediction performance. By contrast, our combined model cleverly combines the strengths of both approaches. It inherits the robust imaging feature extraction capability of the dual-channel model, which enables the detection of subtle features in ultrasound images. Additionally, by incorporating clinical features, it considers individual patient differences and disease progression characteristics, enabling a comprehensive, multi-level analysis of the patient’s condition. The integration of this multidimensional information allows the combined model to more accurately and comprehensively assess the risk of occult CLNM in PTMC patients across both the training and test sets. This significantly enhances the model’s prediction accuracy and reliability, providing clinicians with more precise and effective decision-making support during the diagnosis and treatment of PTMC. It is expected to play a key role in future clinical practice and contribute to the continued advancement of thyroid cancer diagnosis and treatment.

This study has several limitations. First, data were obtained from only two hospital centers, which may limit the generalizability of the findings. Second, the retrospective design introduces potential information bias due to incomplete or inaccurate medical records, complicating the control of confounding factors. Additionally, the AI model is not an end-to-end intelligent system; significant human intervention during the modeling process may introduce subjectivity. The dual-channel model integrates information from only two sections, limiting its ability to capture the full variability of thyroid tumors across multiple sections and volumetric dimensions, which may reduce predictive accuracy. Future research should focus on models incorporating multi-sectional or tumor volume data to better represent tumors and improve predictive efficiency.

The DL model that integrates dual-channel ultrasound imaging information effectively predicts occult CLNM in patients with solitary PTMC.