Functional MRI (fMRI) is an invaluable tool for neuroscience, but it necessitates a trade-off between spatial and temporal resolution to maintain a reasonable temporal signal-to-noise ratio (tSNR). Non-Cartesian acquisition schemes are more efficient sampling strategies as compared to Cartesian ones and were proposed as a way to push the limits of spatiotemporal resolution further. 3D-SPARKLING is a novel non-Cartesian scheme recently evaluated for fMRI applications. However, 3D-SPARKLING is highly sensitive to B0 field imperfections. Such imperfections are detrimental to fMRI applications, especially at ultra-high magnetic fields. In this work, we collect measurements of the static and dynamic B0 field perturbations concurrently with 3D-SPARKLING fMRI data acquisition and retrospectively correct these perturbations during image reconstruction. The advantages of this image reconstruction strategy are assessed on the statistical sensitivity to the BOLD contrast during a retinotopic mapping fMRI experiment. Importantly, a gain of 44% (resp., 159%) additionally activated voxels was quantified when adopting this brain activity-enhanced image reconstruction technique at a type-I statistical control level of 0.001 without multiple comparisons correction (resp., 0.05 with false discovery rate correction). Additionally, significantly improved retinotopic maps were retrieved on the cortical surface.