DGNO Neural Operator Outperforms Existing Methods for Pathology Defocus Deblurring

Defocus deblurring in pathological microscopy has long been hampered by the spatially varying and locally discontinuous nature of optical blur caused by position-dependent integral imaging. Traditional deep learning approaches rely on shift-invariant assumptions and struggle with these heterogeneous blur patterns. Neural operators offered a principled alternative by modeling defocus formation as an integral operator, but most existing architectures use globally parameterized kernels that assume smoothness—limiting their ability to capture real-world discontinuities.

To overcome this, the team proposes the Discontinuous Galerkin Neural Operator (DGNO), which parameterizes the integral kernel using a discontinuous Galerkin formulation with element-local volume operators and interface numerical fluxes. This design combines locality, heterogeneity modeling, and global coherence while preserving the physics of optical image formation. Extensive experiments show DGNO surpasses state-of-the-art methods, delivering significantly sharper reconstructions, robust performance on spatially varying blur, and scalability to high-resolution inputs. Accepted at ICML 2026, the code will be released, making this a strong candidate for medical imaging pipelines.