<p>Zero-watermarking technology emerges as a potent tool for protecting the copyright of medical images. Traditional zero-watermarking schemes often rely on manual feature extraction using conventional mathematical techniques, yet they cannot fully protect medical images. Recognizing the advantages of neural networks for image feature extraction, this paper introduces a robust zero-watermarking scheme based on a denoising convolutional autoencoder (DCAE) to enhance its robustness against medical images. The scheme uses a pretrained DCAE to obtain robust features. Subsequently, the watermark image scrambled by the logistic chaotic map is XOR-operated with the binary feature vector transformed from robust features via the average hash (aHash) algorithm to generate a zero-watermark. The normalized correlation coefficient (NC) values obtained from the robustness experiments in this paper are above 0.8. This indicates that the zero-watermarking scheme proposed in this paper is highly robust and is generally superior to the compared algorithms, providing a guideline for securing copyright and privacy information in medical images.</p>