The widespread adoption of telemedicine has led to an increased digital transmission of medical images. However, in the current complex network environment, these images, which often contain sensitive patient information, are susceptible to malicious attacks, posing a significant threat to patient privacy and information security. To address this issue, we propose a robust zero watermarking algorithm for medical images, leveraging discrete wavelet transform (DWT), convolutional neural networks (CNNs), EfficientNet, and discrete cosine transform (DCT). Firstly, the original medical image is transformed by DWT to extract the LL low-frequency sub-band, which is then fed into the EfficientNet neural network to obtain the global average pooling layer coefficients and perform DCT transform. Further, we process the low frequency of DCT to generate a 32-bit hash sequence and perform XOR operation with the chaotic encrypted watermark to produce the key for watermark embedding. Notably, watermark extraction can be robustly achieved without the original image. Simulation results show that the algorithm has good robustness and strong resistance to geometric attacks.