OTSCC has poor prognosis and high morbidity among OSCC[3, 30]. CLNM is an adverse factor for low survival rate and poor prognosis of OTSCC patients [31–33]. Unfortunately, there was a lack of particularly optimized medical techniques for identifying CLNM for OTSCC patients.
Radiomics can extract high-throughput quantitative radiomics features from radiological images and build disease classification models[13]. There have been some early researches on the application of radiomics in determining CLNM for OTSCC patients. Takaharu Kudoh et al. designed radiomics analysis of [18F]-fluoro-2-deoxyglucose positron emission tomography for the prediction of cervical lymph node metastasis in OTSCC, with the AUC of 0.79, but they just enrolled 40 patients and lacked a testing set[34]. Masaru Konishi et al. constructed radiomics machine learning model based on radiomics analysis of intraoral ultrasound images for prediction of late CLNM for OTSCC patients and achieved excellent performance, with the AUCs of 0.995 and 0.967 in training group and validation group[35]. But they have not further done subgroup analysis for different tumor staging. Yiwei Zhong et al. established artificial network model based on CT radiomics signature, with the AUC of 0.943 (0.891–0.996), but they ignored the effect of clinical risk factors[36]. Previous studies also have confirmed that MRI radiomics can evaluate preoperative CLNM for HNSCC and other sites cancers[20–23, 37–39]. But they mostly were single-center studies or single-sequence MRI studies. In a single-center, Fei Wang et al.[40] reported that the radiomics features model with peritumoral 10mm expansion can predict CLNM in OTSCC just based on T2WI, with the AUCs of 0.995 (0.991–0.999) and 0.872 (0.847–0.897). On the one hand, their study lacked an independent external testing set. On other hand, the proportion of peritumoral 10mm expansion varied depending on the size of the tumor, especially between T1-T2 and T3-T4. It is of great benefit to further explore the additional value of the radiomics features based on multi-modal MRI, and multi-center studies were required to reach more convincing conclusions.
In our multi-center study, we constructed two combined models including two-sequence MRI and three-sequence MRI, respectively. The combined model based on three-sequence MRI had best performance than the other models and it presented highest clinical overall net benefit. Moreover, both radiomics model and combined model in group Ⅱ had better performance than the models in group Ⅰ. This study assumed that the radiomics features of CE-MRI may have better robustness[41]. It captured the tumor intra-heterogeneity, which is because the CE-MRI was less affected by peritumoral inflammatory and edema compared with T1WI and FS-T2WI, and VOI in CE-MRI could reflect the whole tumor issue more precisely[42].
There were a few studies on predicting CLNM for OTSCC patients in cN0 and cT1-T2, and most researchers just built prediction models, but they did not validate them[43, 44]. Pensiri Saenthaveesuk et al.[45] designed a nomogram based on tumor size, tumor location, radiologic tumor thickness threshold and paralingual distance measured in multiparametric MRI to predict OCLNM risk in early OTSCC, with the AUCs of 0.952 (0.917–0.987) in the primary cohort and 0.881 (0.779–0.983) in the validation cohort, but they did not think highly of the superior quantitative radiomics features, and they lacked an external test set..
In this study, we demonstrated that the combined model integrating clinical factors and multi-modal MRI radiomics features on the basis of three-sequence MRI can well predict preoperative CLNM for cN0 and cT1-T2 OTSCC patients, which suggested that it was beneficial for early OTSCC patients to accept CE-MRI examination and that it also provided evidence for the diagnosis and operation of the patients.
The results of this study suggested that the clinical factors could not be overlooked for research CLNM. DOI is closely related to the CLNM, prognosis and decision of neck dissection strategy[46–48]. Our study showed that DOI (OR: 1.131, 95%CI: 1.031–1.253) based on MRI was an independent risk factor for CLNM in OTSCC patients. Kazuya Haraguchi et al.[49] also demonstrated a significant correlation between the DOI of GdT1WI and T2WI and histopathological DOI, and they further determined the cut-off value of radiological DOI, which was 6.99mm in GdT1WI and 8.32mm in T2WI. PNI is an important risk factor affecting the prognosis and CLNM in OSCC patient, and cancers with PNI tend to be more invasive[4, 50]. Tongue pain and dyskinesia reflect the clinical signs of PNI. We took the method given by Dae Young Yoon et al. to minimize the subjectivity of surgeons to discriminate RLNM[51, 52]. Our study also showed that RLNM (OR: 4.378, 95%CI: 1.916–10.476) correlated with CLNM.
Some limitations of the study also needed to be noted. First, this multi-center included three MRI scanners with different parameters, which was a strong confounding factor needed more effective measures to weaken center effect. Second, this study was a retrospective study with selection bias. Third, our models only focused on whether OTSCC patients had CLNM, but the number of positive lymph nodes and contralateral CLNM also affected the prognosis. Finally, in addition to T1WI, FS-T2WI and DCE-MRI, MRI included many other sequences needed to be further explore.