Pretreatment multiparametric MRI radiomics-integrated clinical hematological biomarkers can predict early rapid metastasis in patients with nasopharyngeal carcinoma

Nasopharyngeal carcinoma (NPC) has the highest incidence among malignant tumors of the head and neck [1‐3]. Radiotherapy plays a fundamental role in the treatment of nasopharyngeal carcinoma (NPC) due to its remarkable degree of radiosensitivity. The occurrence of local recurrence and/or distant metastases constitutes the primary cause of therapy failure in patients with this condition. The implementation of intensity-modulated radiation therapy and concomitant chemoradiotherapy has significantly enhanced the local control rate of NPC, leading to a remarkable 5-year overall survival (OS) rate of 80%-88%. However, it is important to note that distant metastasis has emerged as the predominant pattern of treatment failure, accounting for the majority of cases [4]. Approximately 10% of individuals with NPC present with distant metastasis upon initial diagnosis, while an additional 10–20% progress to metastasis following treatment, resulting in a 5-year survival rate below 10% [5]. Metastasis profoundly affects patients' quality of life and rapidly progresses, leading to mortality [6].

In clinical practice, the tumor-node-metastasis (TNM) classification system serves as a crucial tool for clinical decision-making and prognostic assessment in tumor patients, including those with nasopharyngeal carcinoma (NPC). Although patients with the same classification are typically treated with similar therapeutic approaches, their outcomes and prognoses may significantly differ. This discrepancy suggests that the TNM classification system primarily considers the relationship between the tumor and surrounding tissues and organs and fails to incorporate intratumor characteristics and heterogeneity, thereby limiting its ability to accurately predict risk stratification markers. In contrast, magnetic resonance imaging (MRI), with its superior soft-tissue resolution, offers more precise visualization of microscopic lesions and the exact extent of lesions [7]. Currently, MRI is widely employed for tumor staging, image guidance, and follow-up via manual visual interpretation. However, the clinical application of MRI is primarily focused on qualitative structural information. To overcome this limitation, a modern technology called 'radiomics' [8] can extract high-throughput quantitative information, enabling its use in clinical diagnosis, prognosis, and treatment evaluation. Numerous studies across various cancer types, including head and neck squamous cell carcinoma [9‐11], lung cancer [12], rectal cancer [13, 14], and breast cancer [15, 16], have demonstrated the prognostic value of radiomics. These studies have linked radiomics to local recurrence, pathological molecular classification, and pathological remission after neoadjuvant chemotherapy [17]. Additionally, the formation of oxygen-free radicals resulting from persistent oxidative stress and the inflammatory response is directly associated with the onset and progression of cancer from a pathophysiological perspective [18]. Thus, inexpensive and easily identifiable inflammatory blood markers, such as the lymphocyte-to-monocyte ratio (LMR), platelet-to-lymphocyte ratio (PLR), neutrophil-to-lymphocyte ratio (NLR), and systemic immune-inflammation index (SII), have emerged as potential prognostic indicators for NPC [19, 20]. Consequently, there is a crucial need to explore biomarkers accurately from multiple dimensions. This exploration could facilitate the development of more aggressive and personalized treatment plans for high-risk patients in a timely manner. To address this need, we constructed and validated a combined model aimed at early prognostication of the likelihood of rapid metastasis.

In this study, we constructed a combined radiomic and clinical model based on radiomic features from baseline MR images before initiating treatment and clinical variables to predict the risk of early distant metastasis in NPC patients. The combined progression model showed superior predictive ability and was remarkably superior to the clinical model and radiomic model alone. This is the first study in which we extracted signatures of primary tumors from four-sequence MR images and evaluated the feasibility of predicting progression over two years by MR imaging-based radiomic features of local advanced nasopharyngeal carcinoma.

Because the effectiveness of radiotherapy and concurrent radiotherapy and chemotherapy significantly improved the local control rate of NPC, distant metastasis was the main failure mode. First, because approximately 59–73.5% of distant metastases occur within the first 2 years after complete remission of the nasopharynx tumor, early prediction of a high risk of disease progression after complete regression is vital and abstruse. This is an obstacle to adopting early intervention or more aggressive management for high-risk patients. Second, the treatment tolerance and reaction of patients with early distant metastasis are poor after simultaneous chemoradiotherapy, and the 5-year overall survival rate is less than 5% [22]. Therefore, it is necessary to predict the high risk of early metastasis in NPC patients before treatment to optimize individualized therapeutic strategies for early prevention [23‐30]. Chen et al. reported the importance of predicting progression-free survival because early active intervention has been shown to improve mortality [1].

Due to the existence of tumor heterogeneity [31], it is crucial to find a method to more comprehensively and accurately capture microscopic characteristics and differences in tumors. We considered independent clinical factors and radiomics signatures to establish a composite model for predicting the risk of early distant metastasis. Radiomics signatures can reflect the pathophysiological information intrinsic to the tumor and provide abundant information. Bao D et al. evaluated the value of pretreatment MRI radiomics machine learning models in predicting disease progression in nasopharyngeal carcinoma patients who achieved a complete response after treatment and may help to improve clinical decision-making [28]. A previous study reported a machine learning model based on a combination of clinical and radiomic features in a developed cohort (AUC: 0.80) and a validated cohort (AUC: 0.80) that can discriminate 3-year disease progression equally well after primary treatment [31]. Fang ZY et al. reported that a clinical radiomic model exhibited superior prediction ability and accuracy compared to a simple clinical model or radiomic model alone [32, 33]. In another previous study, they developed a progressive radiomic model based on combined contrast-enhanced T1WI texture and TNM classification and achieved an AUC of 0.78, which was better than that of the TNM classification system alone for discriminating 3-year PFS (C-index: 0.761; 95% CI: 0.664 to 0.858), which was in accordance with our findings [34]. In our study, the AUC of our combined model for predicting early rapid metastasis reached 0.936 (95% CI: 0.901–0.971) in the training cohort and 0.796 (95% CI: 0.686–0.905) in the validation cohort, indicating that our model had better stability than the radiomic model alone.

Previous studies have tried their best to analyze the relationships between radiomics features and clinical outcomes, such as overall survival (OS) [35], local recurrence [24], and side effects after radiotherapy [36]. We focused on the risk of early distant metastasis and a worse prognosis. In contrast to studies that simply identified the lesion's largest cross-section, we constructed a 3D VOI from several successive slices of the whole tumor, which could more accurately represent the lesion's heterogeneity [37]. Tumor heterogeneity may be related to tumor angiogenesis, cell proliferation, necrosis, and even different tumor gene phenotypes [38]. Greater tumor heterogeneity is closely connected with poorer prognosis, which could be only associated with intrinsic aggressive biology or therapy resistance [39]. We extracted radiomics features from four sequences to construct a prognostic model that may be more comprehensive considering the characteristics of different MRI sequences and exhibited excellent performance for individual prediction [40]. Wang et al. reported that the radiomics model derived from multiple MR sequences had better predictive capability than that derived from a single MR sequence (p < 0.05) [41]. Moreover, the radiomics data extracted from functional sequences indicated invasive biological features of the tumor, leading to a high possibility of local disease progression. In our study, 12 radiomic features were retained, and these were the most important factors for predicting 2-year disease progression. First-order statistics describe the distribution of voxel intensities within an image region defined by a mask using commonly utilized basic metrics [42]. The gray-level dependence matrix (GLDM) quantifies the characteristics of gray-level dependency in an image, specifically the similarity or dependency of pixel values on their neighboring pixels. This feature can reveal the textural coarseness of tumor tissue [42]. Tumors with coarser textures might indicate greater tissue heterogeneity, which is sometimes associated with poorer treatment response and prognosis. The gray level co-occurrence matrix (GLCM) provides information on image texture, such as contrast and uniformity, related to the arrangement and structural heterogeneity of tumor cells. Similar to GLDM, these textural features may correlate with the biological behavior and prognosis of the tumor. The gray level size zone matrix (GLSZM) is highly relevant to tumor shrinkage and can quantify textural complexity, implying more complex biological characteristics of the tumor. [43, 44]. GLSZM, GLRLM, and GLCM are regional textural features and have been applied to emphasize local heterogeneity information, as the ability to distinguish patients with distinct prognoses has already been confirmed for other tumors [45‐47]. A previous study demonstrated that heterogeneity in MRI distribution within tumors serves as a valuable biomarker for predicting treatment outcomes in patients with NPC [28]. A study also demonstrated that features such as GLCM, GLDM, and GLSZM extracted from PET-CT can predict local recurrence-free survival (LRFS) in nasopharyngeal carcinoma patients. Our study corroborates these findings, showing that these features can predict early rapid metastasis occurring two years posttreatment.

Previous research has demonstrated that these clinical factors might act as significant prognostic indicators for NPC patients [19], although they only had a modest predictive value for enhancing reclassification performance in this study. Lymphocytes play a major role in the immunological response of the host to tumors. Neutrophil granulocytes play an important role in controlling the circulatory angiogenesis of chemokines, growth factors, and proteases [19, 48]. Thus, we retained them as conceivable elements in the nomogram [49]. Previous studies have investigated the relationship between blood inflammation indicators and NPC prognosis, and an NLR ≤ 2.695 is associated with poor OS [19]. However, one study suggested that patients with an NLR ≥ 3 had worse survival [20]. In this study, the NLR was included in the nomogram model, and an NLR ≥ 3.8 indicated a greater probability of metastasis. This may be related to the diversity of the selected patients. The two different results may suggest a deeper mechanism waiting to be explored.

Our study has several limitations. First, due to the retrospective nature of the study, selection bias may be inevitable. The research sample size in this study was relatively small, and the data were derived solely from a single center. To address potential biases inherent in retrospective data analysis, a model was constructed and subsequently internally and externally validated through multicenter collaborations. However, the validation results may have been compromised due to the limited number of patients in the validation group. Therefore, future prospective studies encompassing larger populations, along with further external validation, are warranted to validate the conclusions drawn from this research. Additionally, it is worth noting that different MRI equipment was utilized for inspections, and the scanning parameters were not standardized. Future endeavors should focus on standardizing imaging protocols. Third, our study revealed the good predictive value of radiomic features, but the present study was limited because pretreatment did not include posttreatment or dynamic characteristics. Therefore, a prediction model that further integrates longitudinal data and images may improve the prediction power.

Our study is the first to incorporate clinical features, hematological markers, and radiomic features, and our nomogram model seems to be an effective predictor of NPC outcomes. To improve the prognosis of NPC patients via layered care, a nomogram based on the NLR, T classification, N classification and radiomics for predicting early rapid metastasis in NPC patients may serve as a therapeutic personalized tool. Further research is also required to explore the generalized utility of our model and to translate it into clinical use.