Multimodal Classification of Parotid Tumors – A Proof of Concept

Abstract

Preoperative differentiation between benign and malignant parotid tumors is challenging but can influence treatment and outcome. In this proof-of-concept study, we investigated multimodal AI-based classification combining clinical tabular data and ultrasound images, comparing unimodal and multimodal approaches for binary benign-malignant classification. The study included 594 patients with parotid tumors (449 benign, 145 malignant) after preprocessing. The image-only classifier achieved 0.73 accuracy at image level and 0.75 after patient-level aggregation. The tabular model reached 0.88 accuracy, while the multimodal model maintained this accuracy with a slightly improved weighted F1- score (0.89) and malignant recall of 0.93. These results suggest that image-derived features provided complementary information, primarily improving sensitivity for malignant tumors.