Least-squares migration can improve image resolution and compensate for uneven illumination, but data-domain implementations require repeated modeling and migration. Image-domain methods based on point-spread functions (PSFs) reduce this cost by approximating the local Hessian response, yet their performance depends on the representativeness of local PSFs in structurally complex areas. We propose SPIRE-Net, a self-supervised migration image restoration framework constrained by spatially variant PSFs. The network predicts a restored image patch, which is then filtered by the PSF assigned to the corresponding spatial location. Training minimizes the mismatch between this PSF-filtered prediction and the input image patch without requiring supervised target images. Structure-aware adaptive PSF-paired sampling and a structure-enhanced consistency loss strengthen the constraint in complex structural regions. The Marmousi experiment demonstrates the ability of SPIRE-Net to solve the prescribed PSF-based image-domain inverse problem through self-supervised learning. Field-data comparisons show improved local reflector continuity and event separation relative to image-domain LSM in the examined regions, while preserving the main reflection geometry. These results suggest that spatially variant PSFs can serve as physical constraints for self-supervised migration image restoration.