From Multimodal Data to Clinical Insight: An Explainable Model for Preoperative Salivary Gland Lesion Diagnosis.

OBJECTIVE

To develop and validate a multimodal dual-step support vector machine model (SVM-DualNet) for the preoperative three-class classification of salivary gland lesions (SGLs) to support clinical decision-making.

METHODS

We retrospectively collected clinical, conventional ultrasound (CUS), shear wave elastography (SWE), and radiomics features from 284 patients with SGLs. For malignancy discrimination and pleomorphic adenoma identification, linear SVM models based on different modality combinations were constructed and compared. The best-performing binary models were sequentially combined to form SVM-DualNet. SHapley Additive exPlanations (SHAP) were applied for global and case-level interpretation and incorporated into diagnostic assistance. Clinical utility was evaluated by comparing the junior radiologist's diagnostic performance before and after SHAP assistance and by comparison with the senior radiologist.

RESULTS

In the test cohort, SVM-DualNet achieved a balanced accuracy of 0.76 and a macro F1 score of 0.82 for the three-class classification. The binary models discriminated malignancy and pleomorphic adenoma with AUCs of 0.90 (95% CI: 0.82-0.97) and 0.85 (95% CI: 0.76-0.94), respectively. SHAP-assisted review improved the junior radiologist's balanced accuracy from 0.55 to 0.70 and macro F1 from 0.57 to 0.75, approaching the senior radiologist's performance.

CONCLUSIONS

The model provides reliable preoperative classification of SGLs and can assist clinicians in decision-making.