<p>Preoperative prediction of microvascular invasion (MVI) in patients with hepatocellular carcinoma (HCC) is of critical clinical importance for guiding surgical planning and postoperative adjuvant therapy. Although several deep learning-based fusion methods have been proposed for this task, most fail to fully exploit the complementary information across multi-sequence MRI, resulting in limited predictive performance. To address this limitation, we developed a multi-sequence 3D fusion network (M3DFuseNet) for MVI prediction. The model employs a 3D ResNet-34 backbone to extract multi-scale features from different MRI sequences, which are then processed by a global aware module (GAM) and a local aware module (LAM). The GAM incorporates positional encoding and attention mechanism to effectively capture cross-sequence interactions, while the LAM uses a bottom-up aggregation pathway with adaptive weighting to progressively fuse multi-scale features. The fused features are finally passed to a classifier for MVI prediction. M3DFuseNet was evaluated on a multi-center dataset collected from the Affiliated Hospital of Xuzhou Medical University and the Eastern Hepatobiliary Surgery Hospital. The experimental results showed that the model achieved an accuracy of 0.894 ± 0.016 and an area under the receiver operating characteristic curve (ACC) of 0.917 ± 0.033 on the internal validation set, while the performance on the external test set reached an accuracy of 0.814 and an AUC of 0.829. These findings indicate that M3DFuseNet holds substantial potential for supporting preoperative clinical decision-making in HCC patients.</p>