Background: This study aimed to assess the diagnostic value of a deep learning (DL)-based multimodal approach that combines ultrasound (US) and magnetic resonance imaging (MRI) features in differentiating benign and malignant ovarian tumors, and to develop an intelligent auxiliary diagnostic tool. Methods: A total of 887 patients (665 benign, 222 malignant) with pathologically confirmed ovarian tumors from 2022 to 2024 were retrospectively enrolled. All patients underwent preoperative US and MRI within one week. A dual-channel DL model was constructed: the US branch (ResNet50) extracted 2D features from grayscale color Doppler flow imaging, while the MRI branch (3D ResNeXt101) extracted 3D features from T2-weighted imaging (T2WI), dynamic contrast-enhanced (DCE)-MRI, and apparent diffusion coefficient (ADC) maps. The extracted features were integrated using an attention-based fusion mechanism. With pathology as the gold standard, the diagnostic performance of the proposed model was compared with US alone, MRI alone, and the Assessment of Different NEoplasias in the adneXa (ADNEX) model. Results: In the test cohort (n = 266), the DL model showed a sensitivity of 92.73%, specificity of 98.58%, accuracy of 97.37%, and an area under the curve (AUC) of 0.957, which were significantly higher than those of US (AUC = 0.792, z = 3.92, p < 0.001), MRI (AUC = 0.844, z = 2.76, p = 0.006), and ADNEX (AUC = 0.885, z = 2.07, p = 0.022). Conclusions: The DL-based US-MRI multimodal fusion model significantly enhances the diagnostic accuracy for differentiating benign and malignant ovarian tumors, providing a promising intelligent auxiliary tool for early and precise diagnosis.