An Interpretable Machine Learning Model With Synthetic MRI-Based Habitat Radiomics for Predicting Lymph Node Metastasis in Oral Cancer.

PURPOSE: To develop an interpretable radiomics model based on habitat for predicting lymph node metastasis in oral cancer using Synthetic MRI (SyMRI).

METHODS

A retrospective study of 101 oral cancer patients with pathologically confirmed lymph node status was performed, dividing them into a training cohort (N = 71) and a test cohort (N = 30). All patients underwent MRI, including the MAGiC (SyMRI) sequence. Intra-tumoral volumes of interest (VOIs) were clustered into three spatial habitats using the K-means algorithm. Radiomics models were developed for the entire tumor and clustered intra-tumor regions, based on features from synthetic quantitative T1 and T2 maps. The Shapley Additive Explanations (SHAP) method was used to interpret predictions. A radiomics nomogram was constructed by integrating independent variables, including the radiomics score and MRI-reported status. Predictive performance was evaluated using AUC, DeLong's test, calibration curves, and decision curve analysis (DCA).

RESULTS

The habitat with the highest T1 and a moderately low T2 values within the tumor showed the best predictive performance, compared to whole-tumor radiomics (training: 0.87 vs. 0.82; test: 0.74 vs. 0.62). The radiomics nomogram, combining radiomics features and independent clinical variables, outperformed clinical diagnosis alone in both the training cohort (0.92 vs. 0.80; p = 0.0023) and test cohort (0.84 vs. 0.73; p = 0.0362).

CONCLUSION

The habitat-based radiomics signature, integrating SyMRI-derived quantitative T1 and T2 maps, provides enhanced predictive accuracy for preoperative lymph node metastasis in oral cancer compared to clinical diagnosis.