Hepatocellular carcinoma (HCC) is a representative primary liver cancer with high incidence and mortality. Surgical resection is the first option of treatment, but patients are usually at a high risk of tumor recurrence within the first 2 years, resulting in poor overall outcomes. Thus, it is of great clinical significance to predict HCC early recurrence to improve the survival rate. Existing transformer-based methods typically rely on models pretrained on ImageNet, which may not effectively extract features from medical images. Thus, we propose a self-supervised transformer-based method for HCC early recurrence prediction. We first leverage a vision transformer trained through self-supervised learning to capture more informative semantic features from multimodal MRI. These enhanced features are subsequently fused with a transformer architecture to predict the early recurrence of HCC. Experiments show that our method achieved better results than other methods.