Early detection of Mild Cognitive Impairment (MCI) with FDG-PET is essential for timely Alzheimer’s disease intervention. However, PET image quality is limited by low spatial resolution, partial volume effects, and Poisson noise. Standard enhancement methods, such as Bilateral filtering or Contrast Limited Adaptive Histogram Equalization (CLAHE), can increase contrast but often introduce heavy noise or distort image texture, while deep learning methods may produce hallucinated structures. We propose a fully data-adaptive, non-learned 3D enhancement framework whose output is deterministic for a given input volume, that combines Laplacian-based local contrast modulation with a gradient-gated Difference-of-Gaussians (DoG) detail injector. This hybrid design sharpens anatomical boundaries while keeping noise amplification near unity in uniform regions. The method enhances structure only where true radiotracer gradients are present. We evaluated the approach on a large Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort ( $N=1928$). Quantitative results show that the method increases contrast without adding noise, achieving a Noise Gain of $1.01$ (vs. $1.28$ for Bilateral filtering) and a high Edge Preservation Index ( $0.981$). In downstream classification experiments across multiple deep learning architectures, the greatest improvement was observed in MobileNetV4 on the axial plane, where mean accuracy increased from 93% to 96%. Overall, the proposed gradient-gated hybrid enhancement provides a reliable PET pre-processing strategy. By recovering subtle metabolic patterns without amplifying noise, it strengthens the sensitivity of automated MCI diagnostic systems.