An intelligent gradient-guided hybrid inpainting framework for brain MRI reconstruction and Alzheimer's disease classification in connected healthcare systems

Brain magnetic resonance imaging (MRI) is essential for early Alzheimer's disease diagnosis, yet clinical scans are often degraded by motion artifacts, signal loss, or incomplete acquisitions. Image inpainting offers a promising preprocessing solution, but existing methods have limitations: deep learning models such as LaMa generate visually plausible reconstructions but may compromise structural fidelity, while classical diffusion-based approaches like OpenCV Telea preserve local continuity but tend to oversmooth complex anatomy. This study proposes a gradient-guided hybrid inpainting framework that integrates OpenCV Telea and LaMa to leverage their complementary strengths. A gradient magnitude-based weighting mechanism enables structure-aware reconstruction, assigning edge-rich regions to the classical method and smoother regions to the deep model, thereby preserving both fine anatomical details and global consistency. Experiments on a four-class ADNI-based MRI dataset, balanced using SMOTE–NM and split 70:15:15, with 10%–30% random masking, demonstrate improved reconstruction performance. The proposed method reduces mean squared error by 8% compared to LaMa and 30% compared to OpenCV, while achieving SSIM ≈0.93 and PSNR ≈25.7 dB. A VGG16 classifier trained on clean images achieves 94.35% accuracy on hybrid-inpainted data, showing only a 1.69 percentage point drop from the baseline and outperforming individual methods. These results highlight the effectiveness of the proposed framework for reliable, AI-driven neuroimaging pipelines in intelligent and connected healthcare systems.