Carefully annotated and processed high-quality datasets are crucial for the training of deep learning models. The commercialization of dataset sales has led to instances of unauthorized dataset disclosure for illicit gains, especially in the medical field, where dataset annotation requires specialized knowledge. Consequently, the verification of dataset copyrights is receiving attention to protect the interests of owners. The multi-modal fusion of text and image is a technology worth studying, in which exclusive information is hidden in the image to achieve the purpose of copyright protection. In the medical field, the fused image needs to meet the requirements of low distortion and text information extraction after being attacked. Slight fusion strength has little damage to the image, but the ability to resist attack is poor and has difficulty in extracting text information from the image. Therefore, an appropriate embedding strength is crucial for multi-modal fusion of text and image to balance imperceptibility and robustness. In this paper, we present an optimization-based approach for obtaining optimal embedding strength. The embedding strength based on the watermark interval is constructed as a differentiable problem, and the loss function is designed to achieve high imperceptibility and strong robustness. In addition, in order to decrease image distortion, an innovative coefficient adjustment scheme based on adjacent blocks is proposed, which distributes the embedding strength into two image blocks. Numerous experimental results indicate the proposed approach offers good imperceptibility and strong robustness when fusing larger text capacities.