<p>Positron emission tomography (PET) provides important functional information for the diagnosis of neurodegenerative diseases; however, its high cost and radiation exposure restrict routine clinical application. Magnetic resonance imaging (MRI) is widely available but mainly captures structural alterations without directly reflecting metabolic activity. Cross-modal translation from MRI to PET therefore offers a practical strategy to bridge this structural–functional gap. Existing MRI-to-PET translation methods, including generative adversarial learning-based, transformer-based, and diffusion-based approaches, primarily focus on structural reconstruction while insufficiently modeling pathology-related metabolic patterns and contrast variations during the generation process. Based on the motivations, we propose CIPHER—a pathology-aware conditional diffusion framework for accurate structure- and pathology-aware MRI-to-PET translation. CIPHER introduces a cross-modal contrast fusion modulator to align multi-contrast representations across modalities and incorporates a temporal consistency constraint to stabilize feature learning across diffusion time steps. This design enhances structural–functional correspondence and improves the preservation of disease-relevant metabolic patterns. Experiments on the ADNI dataset demonstrate that CIPHER achieves improved performance in terms of MAE, MSE, PSNR, and SSIM compared with state-of-the-art baseline methods. These results highlight the structure efficiency of CIPHER and underscore its potential for cross-modality medical image translation.</p>