Image model plays a critical role in recovering diagnosis-relevant information from noisy observation data. Unlike conventional denoising techniques based on local models, a patch-based nonlocal image model is presented and its applications into restoring medical images are demonstrated. We introduce geometric resampling techniques for obtaining redundancy representations which facilitate the exploitation of nonlinear manifold constraint in the patch space. We extend existing locally linear embedding (LLE) into locally linear transform (LLT) to impose sparsity constraint on the uncorrupted images. A nonlocal denoising algorithm based LLT thresholding and adaptive fusion is proposed for removing Rician noise from MRI data and speckle noise from ultrasound images. Encouraging experimental results are achieved, which confirms the value of nonlocal processing as a supplementary tool.