To address the poor applicability of existing reversible adversarial example generation methods in medical imaging, a region selection-based reversible adversarial example generation method for medical images was proposed. Specifically, a generative adversarial network was used to generate precise and subtle adversarial perturbations. Combined with a region segmentation module, these perturbations were selectively constrained to regions that have no impact on diagnostic results. This ensured that the medical images retain high attack performance while maintaining clinical usability. Furthermore, to enable lossless recovery of the original images, a reversible data hiding algorithm was employed to embed the adversarial perturbations in a reversible manner, allowing authorized institutions to legally and securely recover the original data. To prevent overflow issues during reversible embedding, lossless compression was applied to the perturbations before embedding, reducing their size and minimizing the degradation of both attack effectiveness and visual quality caused by the embedding process. Experimental results demonstrate that the proposed method achieves an average attack success rate of 89.27%, a peak signal-to-noise ratio of 29.54 dB, and a structural similarity index of 0.7922, outperforming several classical reversible adversarial example generation methods in overall performance.