Could a multimodal fusion model integrating CT radiomics and systemic inflammatory markers improve preoperative risk stratification of parotid masses? A retrospective exploratory study.

OBJECTIVES

To develop and internally validate a multimodal fusion model integrating clinical features, lymphocyte-to-monocyte ratio (LMR), and venous-phase CT radiomic data for preoperative benign-versus-malignant risk stratification of parotid masses (PMs), and to exploratorily evaluate its potential adjunctive role relative to standard diagnostic modalities including fine-needle aspiration biopsy (FNAB).

METHODS

A retrospective analysis of 490 patients with histopathologically confirmed PMs (2013-2022) was performed, with random assignment to training (70%) and internal validation (30%) cohorts. Radiomic features from venous-phase CT were selected via t-tests, mRMR, and LASSO regression, with all feature selection performed strictly within training folds. Eleven machine learning models were trained and optimized by 10-fold cross-validation, and the model with the highest cross-validated AUC was selected as the optimal radiomic signature. The previously established LMR-based clinical-laboratory signature was re-validated, and logistic regression was used to fuse the two signatures into a multimodal model. Model performance was assessed using receiver operating characteristic analysis, Hosmer-Lemeshow test, calibration curves, and decision curve analysis (DCA). As an exploratory secondary analysis, in 135 patients with preoperative FNAB, the model's risk estimates were compared against retrospectively assigned MSRSGC categories using DeLong's test, net reclassification improvement (NRI), and integrated discrimination improvement (IDI).

RESULTS

The multimodal model achieved an AUC of 0.931 (95% CI: 0.888-0.973) in the internal validation set, outperforming the clinical-laboratory signature (AUC = 0.761; p < 0.001) and radiomic signature (AUC = 0.919; p = 0.038), with sensitivity of 0.957, specificity of 0.782, PPV of 0.449, and NPV of 0.990. In an exploratory secondary analysis of the FNAB subset, the model achieved an AUC of 0.952 (95% CI: 0.918-0.987), with higher statistical discrimination than retrospectively assigned MSRSGC categories (AUC = 0.712; p < 0.001; categorical NRI = 0.399, IDI = 0.386, both p < 0.001).

CONCLUSION

The multimodal fusion model demonstrated improved statistical risk discrimination over single-modal approaches in this retrospective cohort. The model is an investigational adjunctive risk-stratification instrument that is not intended to replace FNAB, MRI, or histopathologic diagnosis. These preliminary findings require prospective multicenter validation before clinical consideration.