Introduction
Lung cancer is the leading cause of cancer-related mortality worldwide; among its histologic subtypes, non-small cell lung cancer (NSCLC) is predominant [
1,
2]. Although the widespread use of low-dose computed tomography for lung cancer screening is detecting early-stage NSCLC in more patients, more than half of patients are at an advanced stage when diagnosed [
3‐
6]. In advanced NSCLC, the cervical lymph nodes (LNs) are common sites for distant metastasis [
7]. Knowing the status of the cervical LNs is crucial for clinicians to make decisions for patients with NSCLC. According to the 8th TNM classification of lung cancer, NSCLC patients with metastasis in the lower and upper cervical LNs are categorized as stage IIIB and IV, respectively. The recommended first-line treatments for these patients are concurrent chemoradiotherapy, targeted therapy, or immunotherapy, while surgical intervention is inappropriate [
8‐
10]. Therefore, identifying the status of the cervical LNs plays an integral role in accurate pretreatment staging and clinical decision-making for patients with NSCLC.
Neck ultrasound and ultrasound-guided biopsy are commonly used to identify the status of cervical LNs [
11]. Although ultrasound is the preferred method for examining cervical lymph node metastasis (LNM), its diagnostic accuracy can be influenced by various factors. There is still room to improve the performance of ultrasound in diagnosing cervical LNM in NSCLC [
12]. In patients with NSCLC, when suspected cervical LNM is detected by ultrasound, ultrasound-guided biopsy is recommended for further clinical investigation. Although ultrasound-guided biopsy is considered the gold standard for assessing the status of cervical LNs, it is limited by sampling errors and potential complications [
11]. Therefore, a precise and noninvasive diagnostic approach is warranted to evaluate the status of cervical LNs.
Radiomics is a promising approach that utilizes quantitative features extracted from medical images to develop models aimed at supporting clinical decision-making [
13]. Due to the high dimensionality of radiomic features, powerful analytical methods and tools are needed. As a vital branch of artificial intelligence, machine learning algorithms have the potential to enhance the performance of radiomics models [
14]. However, there have been no studies applying ultrasound radiomics based on machine learning to the diagnosis of cervical LNM in NSCLC. We designed the present study to investigate the performance of models based on ultrasound radiomic features and/or descriptive semantic features in diagnosing cervical LNM in patients with NSCLC from three institutes. Additionally, we explored the ability of the machine learning algorithms to optimize model performance.
Discussion
In this study, we constructed LR and RF models using radiomic and/or descriptive semantic features to diagnose cervical LNM in patients with NSCLC. The findings are encouraging and expected to yield novel ideas for the noninvasive and precise diagnosis of cervical LNM. Three major findings were observed. First, ultrasound descriptive semantic features, including the long diameter, shape, and corticomedullary boundary, were independent risk factors for cervical LNM in NSCLC. Second, the radiomics model exhibited superior performance relative to the semantic model in diagnosing cervical LNM, and combining radiomic and descriptive semantic features yielded better performance than the single models. Third, the RF algorithm outperformed the LR algorithm in the development of models diagnosing cervical LNM in NSCLC.
Ultrasound is a noninvasive method for diagnosing cervical LN status in patients with NSCLC [
8]. Ultrasound descriptive semantic features observed by radiologists are widely used in current clinical practice to identify the status of cervical LNs. Normal cervical LNs exhibit a flat or kidney-bean-shaped morphology and a hilum rich in fat [
20]. In contrast, metastatic LNs display a rounded shape and an indistinct boundary on ultrasound imaging [
21]. Our research found that a larger long diameter (average, 22.9 mm), irregular shape, and unclear corticomedullary boundary were independent risk factors for cervical LNM in NSCLC. In another study, the long diameter was identified as a critical risk factor for cervical LNM in nasopharyngeal carcinoma, with an average of 23 mm [
22]. These results suggest that a 23 mm long diameter may be a suitable cutoff value for metastatic cervical LNs, but further research with a larger sample size is needed to validate this finding. The short diameter, perpendicular to the long diameter, remains controversial as a marker identifying cervical LNM because it varies depending on location and patient sex [
23]. In the current study, multivariate analysis showed that it was not significantly associated with the cervical LNM in NSCLC. Defined as the ratio between the short and long diameter (SD/LD) of the node, the shape index clinically indicates malignant LNs when it is greater than 0.5, particularly metastatic LNs [
21]. The shape index in this study was visually evaluated by observing the SD/LD and boundary of the cervical LNs. Our results provide evidence in support of the idea that the shape index is applicable in the identification of cervical LNM in NSCLC. In addition, an unclear corticomedullary boundary caused by uneven thickening suggests LNM, and our study demonstrated its association with cervical LNM in NSCLC. Nonetheless, relying solely on a single ultrasound descriptive semantic feature may prove inadequate in differentiating between metastatic and nonmetastatic cervical LNs [
23]. Integrating crucial descriptive semantic features might provide more accurate differentiation. Therefore, we constructed a semantic model using features including long diameter, shape, and corticomedullary boundary, which had been selected through multivariate analysis. Our results demonstrated that the semantic model performed well in distinguishing between patients with NSCLC who had or did not have cervical LNM. Nevertheless, it should be noted that these descriptive semantic features are subjective and rely on the clinical expertise of the radiologists.
Radiomics can be used to extract quantitative features imperceptible to the naked eye from medical images, reflecting physiological, pathological, and genetic information in tumors [
24,
25]. Although the use of radiomics in ultrasound is less common than in magnetic resonance imaging and computed tomography, an increasing number of studies have demonstrated the considerable potential of ultrasound-based radiomics for disease diagnosis and treatment [
26,
27]. Using ultrasound-based radiomics, Zheng and her colleagues [
27] reported a model that could predict the metastatic extent of the axillary lymph node in early-stage breast cancer. Wen and his colleagues [
26] found that the model based on radiomic features outperformed the clinical model using independent clinical risk factors in predicting central cervical LNM in papillary thyroid carcinoma. Despite promising results in diagnosing cervical LN diseases, no studies have investigated the use of ultrasound radiomics for diagnosing cervical LNM in NSCLC until now. The current study represents the first project to employ ultrasound radiomics to diagnose cervical LNM in NSCLC, providing a novel noninvasive approach for clinical diagnosis. Consistent with previous studies, the fivefold cross-validated average AUC demonstrated the superiority of the radiomics model over the semantic model in diagnosing cervical LNM in NSCLC. Our results suggest that ultrasound radiomic features contain valuable information for diagnosing cervical LNM in patients with NSCLC.
Although our study demonstrated the excellent performance of the radiomics model, the clinical utility of descriptive semantic features should not be ignored. Combining the two types of features was shown to improve the performance of the individual models in this study. Min and his colleagues [
22] presented a model that integrated radiomic and descriptive semantic features, achieving better performance than individual models in discriminating between benign and metastatic cervical LNs in patients with nasopharyngeal carcinoma. Consistent with their findings, our study found that the semantic-radiomics combined model performed better than individual models. These results highlight the added value of a combined approach in leveraging diverse information sources to achieve more accurate and robust classification. However, Min and his colleagues employed only the conventional LR algorithm when constructing their model, and applying other machine learning algorithms may improve its performance.
Numerous studies have highlighted the strong capacity of machine learning algorithms to develop prediction and classification models [
28,
29]. In our previous study, we developed RF models based on radiomic features to classify thymomas and thymic carcinomas and distinguish early and advanced TNM stages of thymic epithelial tumors, with satisfactory performance [
25]. Thus, we also employed the RF algorithm to construct models in the current study. The RF algorithm generates multiple decision trees and outputs the classification representing the predominant mode of the constituent trees during training. The ability of the RF algorithm to capture nonlinear interactions in the data makes it helpful in addressing complex and nonlinear relationships between variables. In contrast, the LR performs worse than the RF in analyzing nonlinear relationships and is easily affected by extreme values [
30]. Our result is consistent with previous studies, as the RF model outperformed the LR model, emphasizing the robustness and effectiveness of the RF algorithm in the context of complex radiomic features. Moreover, the MDA was utilized to evaluate the performance of each feature, enabling researchers to focus on those features with a more substantial impact on the overall performance of the model. The original_shape2D_Elongatios was a significant modeling feature with the highest MDA value. Calculated as the ratio of the maximum length to the minimum length in the ROI shape, it underscores the critical role of the shape index in diagnosing cervical LNM in NSCLC. Additionally, 9 out of 15 features included in the RF model were wavelet features, accounting for the majority of the modeling features. This finding aligns with prior research incorporating wavelet features into radiomics models [
31,
32]. The possible reason is that wavelet features may reflect spatial heterogeneity at multiple scales within tumor regions, but further research is necessary to investigate their correlation with pathological information.
Our study has some limitations. First, this is a retrospective study, which may have resulted in selection bias. Although three institutions were included in this study, we did not divide the external dataset in the main text accordingly due to the limited sample size. The significance of the differences between models was not tested because it tends to be not statistically significant with a small sample size. Therefore, a prospective study with a large sample size would be necessary to generalize our findings. Second, it is important to note that most cervical LNs that underwent ultrasound and ultrasound-guided biopsy examinations were suspected of metastasis based on palpable enlargement in this study. This led to a higher rate of cervical LNM in the sample, which is unusual and not representative for general NSCLC cohorts. In addition, only 37 of 313 patients were negative cases, which may have influenced the robustness of the models. To mitigate this issue, increasing the number of negative cases through targeted recruitment or data augmentation techniques may be considered. Third, the patients from the three institutions were examined with different ultrasound devices, which may affect the reproducibility and reliability of the radiomic features. Unified and standardized acquisition and reconstruction parameters may help mitigate this problem. Fourth, we used conventional ultrasound only, so we could not evaluate blood flow and other parameters. Contrast-enhanced ultrasound can display tumor vascularization, while shear wave elastography can assess tissue hardness. Thus, incorporating multimodal ultrasound radiomics analysis may improve the final performance when differentiating between cervical LNM-positive and cervical LNM-negative groups. In the future, prospective studies with larger sample sizes are needed to improve the performance of the models for diagnosing cervical LNM in patients with NSCLC.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.