To the best of our knowledge, there are only few studies illustrating whether an exclusive morphology-based radiomics analysis would enable the assessment of LNM and ENE risks in OTSCC.
The current study constructed machine-learning models with clinical and radiomics morphological features extracted from preoperative MRI to predict lymph node status in OTSCC. For predicting LNM, the NN model showed the best diagnostic performance with an AUC of 0.746 and accuracy of 72.2%. For predicting ENE, the SVM model showed the best diagnostic performance with an AUC of 0.750 and accuracy of 85.6%.
Radiomics is a process of high-throughput extraction of quantitative features that result in the conversion of images into mineable data and the subsequent analysis of these data for clinical decision support [17]. Previous research showed that radiomics could non-invasively help in predicting cervical lymph node status in oral cancers, thus showing tremendous potential for clinical diagnosis and treatment decisions [12, 18]. However, as most previous studies have focused on histogram and texture analysis, the literature on shape analysis is relatively sparse. Radiomics-derived shape features capture the geometric properties of ROIs which are not directly based on voxel values, and should therefore not be greatly affected by interpolation effects [19]. Moreover, shape-based features are conceptually much simpler than other radiomic features. Wang et al. [20] developed a MRI-based machine learning model to evaluate cervical lymph node status in head and neck cancers, and four morphologic features were included in the model but without detailed description. Kubo et al. [11] selected Maximum2DDiameter as a shape feature into the model after the least absolute shrinkage and selection of operator logistic regression analysis. However, their contouring was performed at each neck node level instead of the primary tumor. In our study, we extracted the radiomics-based morphological features from the primary tumor and developed morphology-based machine-learning models in patients with OTSCC. Notably, our customized metric—SideArea—was not selected in the models, which may be related to the resolution of the axial plane; therefore, the estimation of its contact area was inaccurate. Another customized metric—TopBottomArea—was included in all models, which might reflect the extent of tumor invasion from a more comprehensive perspective. Further investigation of the predictive performance of TopBottomArea is warranted in future studies.
Machine-learning models incorporating MRI-derived features from primary tumors can help to identify LNM in several tumors, with reported AUCs ranging from 0.79 to 0.90 [21, 22]. Shan et al. [13] confirmed that machine learning combined with simple clinical and pathologic features showed a better performance in predicting lymph node metastasis of early-stage OTSCC than conventional prediction methods (AUC 0.786 vs. 0.539). However, their model included both preoperative and intraoperative information, which limited its predictive application before surgery. In the current study, all clinical characteristics and radiomics features were from preoperative data, thereby helping in surgical planning.
The presence of ENE places patients in a higher stage, i.e., to either N2a for single small nodal metastases or N3b for multiple or large nodal metastases [14]. Identifying ENE by clinical and radiological examination is difficult, thereby leading to unnecessary overtreatment of the neck by surgical or chemoradiotherapy interventions. Currently, no definitive predictors are available for ENE [23]. Frood et al. [24] found that MRI textural analysis may aid in predicting ENE in oral cancers, with an accuracy of 79%. In the present study, by using the SVM model based on shape features from T2WI and clinical characteristics from ceT1WI, 85.6% of the nodes could be correctly classified preoperatively. This suggests that morphological features combined with machine-learning methods can provide more information and may be a better method for identifying ENE in OTSCC.
We used the IG algorithm to select potential indicators and constructed the machine-learning models. This method allowed us to incorporate individual radiomics morphological features into a feature panel to perform multi-feature analyses. The present study results indicate that the combination of extracted radiomics and clinical traditional morphological variables has a complementary and synergistic effect in predicting LNM and ENE in patients with OTSCC.
DOI was an important independent prognostic factor for lymph node metastasis and survival in patients with oral cancer. For every 5 mm increase in DOI, the T category increases by a level. Previous studies have indicated that MRI-derived DOI had high agreement with pathological DOI and was an independent prognostic factor for LNM in OTSCC [7, 25]. Consistently, the current study identified MRI-derived DOI as statistically significant between LNM-positive and LNM-negative groups. Ren et al. [18] reported that MRI-derived DOI yielded an AUC of 0.67 for predicting occult cervical LNM in early-stage OTSCC. Similar to their study, our results showed an AUC of 0.655 when using MRI-derived DOI alone. Although the sensitivity of MRI-derived DOI was very high, with being 91.1% at the optimal cut-off, the specificity was only 42.2%. Therefore, MRI-derived DOI alone cannot be a reliable indicator to predict the occurrence of LNM. Our study indicated that radiomics shape features potentially improve the ability of the MRI-derived DOI to predict LNM.
There are several limitations in the study. First, this was a retrospective single-institution design without an external validation. Though a stratified 10-fold cross-validation was used to preliminarily verify the feasibility of the study, a prospective multicenter study would be needed for further validation of the models and reduce the risk of overfitting. Secondly, the sample size for analysis was relatively small, especially the proportion of ENE patients. Such imbalance might influence the application of our machine learning models. Future studies should be conducted on a larger cohort, so that we can obtain adequate cases to further train the models and explore the relationship of lymph node status and morphological features. Thirdly, tumors were manually segmented, which may have introduced potential biases, so future research needs a reliable or widely accepted automated segmentation technique to extract the morphological and radiomic parameters of tumors.