<p>Nowadays, medical images are essential for clinical diagnosis and telemedicine. To protect the private medical image information from being illegally accessed or misused, a new visually secure medical image encryption algorithm is raised based on an improved You Only Look Once (YOLO) v11 model and a region of interest (ROI) replacement strategy. The improved YOLOv11 is established by integrating a dynamic feature fusion module with a multi-scale dilated attention mechanism, which enables adaptive fusion of multi-scale local feature maps and bolsters spatial context modeling capabilities. Experimental results reveal that the improved YOLOv11 model demonstrates superior lesion detection accuracy. In our medical image encryption algorithm, the improved YOLOv11 is employed to detect lesion ROIs of the plaintext image. The extracted ROIs are then merged and subjected to multi-level bit-plane decomposition, followed by independent scrambling and diffusion operations at each level to generate an intermediate encryption ROI image. To resist the attack exploiting image differences to infer ROI locations, an ROI replacement strategy is designed, which adopts an image inpainting model to synthesize natural-looking pseudo-ROIs to replace the ROIs in the plaintext image. The encryption ROI image is embedded into the constructed pseudo-plaintext image with a matrix encoding method, which ensures the visual security of the real data. Simulation results verify the feasibility and the effectiveness of our suggested medical image encryption algorithm. For the encryption ROI images, the chi-square values are below 284.3359, the correlation coefficients approach 0, and the information entropies are near 8 bits. The average PSNR values of the ciphertext images and the decryption ones are 51.7619 dB and 51.8594 dB, respectively.</p>