Liver cancer remains a significant global health concern, and early-stage differential diagnosis and treatment can dramatically improve patient survival rates. However, current diagnostic methods are challenging and subjective, highlighting the need for accurate diagnostic support tools. This paper introduces a deep learning-based diagnostic framework for detecting liver tumor malignancy using DCE-MRI, particularly in scenarios with limited well-annotated data. To address the overfitting issue, we propose a Siamese convolutional neural network incorporating metric learning concepts, emphasizing discriminative representation learning. The core principle of our framework is to reorganize the embedding space such that similar sample pairs are closer together while dissimilar ones are further apart. Once the embedding space is learned, unseen samples can be classified based on a distance metric to a few seen samples within the embedding space. Comprehensive experiments on a professionally annotated, self-established dataset demonstrate that our framework outperforms fully supervised approaches in data scarcity situations, achieving a sensitivity of 96%, precision of 82%, and an F1-score of 88%.