In-vivo submillimeter resolution diffusion MRI suffers from limited signal-to-noise ratio (SNR) due to the small voxel size. Denoising techniques can improve the SNR and facilitate further dMRI analysis. Among them, perhaps PCA-based (e.g, Marchenko-Pastur PCA) have shown the best performance. In this work, we introduce kernel PCA, a powerful nonlinear generalization of linear PCA to Hilbert spaces that is shown to suppress a substantial amount of noise (which MP-PCA is incapable of) and still reliably preserve dMRI signal. We showcase K-PCA noise removal with 660 micrometer gSlider data, where we compared it qualitatively and qualitatively with MP-PCA.