To address the limitations of traditional DCT-based medical image watermarking algorithms, particularly their poor robustness under geometric attacks, this paper proposes a novel deep learning-based robust zero-watermarking algorithm for medical images. The proposed method leverages ResNet-50, a deep convolutional neural network, to enhance feature extraction and robustness. The network utilizes low-frequency features obtained from the discrete cosine transform (DCT) of medical images as labels. By incorporating skip connections and a novel objective function, the model strengthens the extraction of high-level semantic features, which are capable of effectively distinguishing between different medical images. These features are then binarized to generate robust hash vectors, which are subsequently bound with a chaotically encrypted watermark to produce corresponding keys, completing the watermark generation process. Notably, the proposed algorithm does not modify the original medical image during the watermark generation stage, nor does it require the original image during the watermark extraction stage. Additionally, the algorithm is designed to support multiple watermarks, further enhancing its versatility. Experimental results demonstrate that the proposed algorithm exhibits strong robustness against both conventional and geometric attacks, with the Rotation attack delivering the best overall performance. This is evidenced by the highest similarity (avgNCC \(\approx \) 0.9995), owest error rate (avgBER \(\approx \) 0.0004), and best image quality preservation (avgPSNR \(\approx \) 30.744), showcasing exceptional robustness and reliability. The proposed approach outperforms existing methods in terms of reliability and security, providing a significant advancement in medical image watermarking. It ensures the protection of sensitive data while maintaining image integrity, making it a promising solution for secure medical image sharing and storage.