Abstract: Classification of liver tumors into benign and malignant is an important research direction in computer-aided diagnosis for medical imaging. However, dynamic contrast-enhanced ultrasound (CEUS) images are characterized by long temporal sequences, sparse key events, and small lesion regions, making it challenging for existing methods to effectively model their spatiotemporal features. To address these limitations, this paper proposes a spatiotemporal feature modeling-based method for benign and malignant classification of liver tumors in dynamic CEUS images. First, a temporal deformable attention module is introduced to adaptively select key reference frames and enhance temporal feature extraction. Second, a spatial attention mechanism is employed to strengthen the representation of lesion regions and their surrounding tissues. Finally, a temporal attention mechanism assigns adaptive weights to individual frames, enabling the model to focus on the most informative frames for classification. Experimental results demonstrate that the proposed model achieves an accuracy of 84.2%, an F1-score of 84.1%, and an AUROC of 90.9%, outperforming conventional deep learning approaches and other temporal modeling methods. Ablation studies and statistical analyses further verify the effectiveness of the proposed modules. The proposed method achieves high classification accuracy and computational efficiency, providing a promising and interpretable auxiliary solution for the automatic diagnosis of liver tumors.