- The paper introduces DGNO as an operator-learning framework that decomposes blur inversion into element-local operators and learnable interface fluxes.
- The method achieves superior PSNR and SSIM on standard microscopy datasets by effectively handling spatially varying, discontinuous blur.
- DGNO’s architecture integrates multi-scale feature extraction and stable operator inversion, enhancing downstream tasks like cell segmentation.
Discontinuous Galerkin Neural Operator for Pathology Defocus Deblurring
Introduction and Problem Context
Defocus deblurring in pathological microscopy is a longstanding, unresolved problem characterized by spatially varying and locally discontinuous optical blur. Existing image restoration models—including CNNs, vision transformers, and state-space architectures—demonstrate limited effectiveness when applied to heterogeneous defocus degradations typical in complex tissue imaging. These approaches universally inherit assumptions such as shift-invariance (CNNs), oversmoothing via excessive global coupling (transformers), or a lack of physical interpretability (state-space models like Mamba). The discrete, pixel-based regression formulation underpinning these models fails to leverage the inherent operator-theoretic structure of the physical image formation process.
The spatially varying point-spread function (PSF) in real optical microscopy is a position-dependent operator, making the underlying blur formation a non-stationary integral operator problem. The inverse problem—recovering a sharp image from a blurred observation—thus demands an operator learning perspective, where the model inverts a highly heterogeneous, piecewise continuous mapping between infinite-dimensional function spaces.
Proposed Method: Discontinuous Galerkin Neural Operator (DGNO)
The Discontinuous Galerkin Neural Operator (DGNO) introduces a principled operator-learning framework for the pathological defocus deblurring problem, leveraging techniques from the discontinuous Galerkin (DG) paradigm traditionally used in the numerical analysis of PDEs. Unlike classical neural operators that rely on globally parameterized kernels and assume stationarity or smoothness, DGNO decomposes the global operator into a sum of element-local volume operators and interface numerical flux terms.
Element-local operators model spatial locality and preserve piecewise structures typical of microscopic images with tissue heterogeneity, while interface fluxes (central, upwind, jump, average-jump) provide explicit, learnable control over information exchange across element boundaries, maintaining global coherence and stabilizing the operator inversion. The DGNO architecture supports both interface-based (DGNO-Face) and zero-order (P0DG, DGNO-Cell) interface formulations, with the latter integrating interface effects via neighboring cellwise representations.
DGNO integrates multi-scale hierarchical feature extraction, instance-specific attention-based basis learning, and a minimal but effective number (T=2) of iterative kernel integral operator layers, ensuring computational scalability to high-resolution imagery. Network parameters—including encoder channelization and window dimensions—are tuned for robustness to microscopy image statistics.
Experimental Results and Quantitative Evaluation
Comprehensive evaluation across standard microscopy blur datasets (BBBC006w1, BBBC006w2, 3DHistech) demonstrates that DGNO yields the highest PSNR and SSIM values compared to all contemporary deep learning baselines, including NAFNet, Restormer, MambalRv2, and MPT+EFCR. On the challenging BBBC006w1 dataset, DGNO achieves up to 37.22 dB PSNR, outperforming state-of-the-art transformer and operator-based architectures by margins up to 1.87 dB, with DGNO-Face and DGNO-Cell providing complementary strengths depending on blur complexity.
Ablation studies against the global Galerkin operator baseline (SRNO), windowed Galerkin variants, and different flux/boundary condition configurations confirm that the introduction of DG-style interface terms is statistically significant for heterogeneous blur modeling. DGNO achieves superior effective-rank utilization of latent representations—allowing the model to learn richer, more discontinuous basis patterns—without incurring computational overhead.
Edge-aware PSNR analysis and visualization of error distributions reveal DGNO's marked advantage in restoring fine cellular structures and sharp boundaries, surpassing global coupling models which tend toward oversmoothing at transitions. Controlled synthetic experiments with spatially varying Gaussian kernels and real-world settings (using known spatially varying PSFs) further validate the robustness and generalizability of the learned operator.
DGNO also excels on natural image datasets (DPDD, RealDOF), outperforming leading SOTA models, indicating that the underlying operator-based formulation is not restricted to pathology. Downstream analysis on cell detection with StarDist affirms that restoration by DGNO directly enhances the accuracy of subsequent segmentation and phenotyping tasks, nearly matching the performance seen on ground-truth sharp images.
Theoretical and Practical Implications
The DGNO framework bridges the gap between operator learning and physically informed image modeling by enabling nonstationary, locally adaptive, yet globally stable neural operator inversion. This harmonizes the opposing demands of spatial locality (crucial for heterogeneous biological specimens) and cross-domain coherence (to avoid artifacts and preserve macrostructure).
From a computational perspective, DGNO attains favorable runtime and memory scaling relative to monolithic global operators and attention-based transformer models, attributable to its elementwise decomposition and controlled message passing via learnable numerical fluxes. The flexibility to instantiate boundary conditions (Dirichlet, Neumann, periodic) directly at the operator level further enhances adaptability to diverse image acquisition protocols.
In principle, this approach suggests new avenues for interpretable, operator-theoretic modeling in other imaging modalities plagued by nonstationary, piecewise-degraded signals. The methodology's modularity and compatibility with existing encoder architectures suggest straightforward integration into more general computer vision or biomedical analysis pipelines.
Future Directions
Ongoing developments may include:
- Exploring higher-order DGNO variants for more complex interface interactions,
- Automated penalty parameter selection for numerical flux terms,
- Adapting DGNO for video and 3D volumetric restoration,
- Theoretical analysis of operator invertibility and generalization in the infinite-dimensional limit,
- Extensions to other spatially varying inverse problems, including multi-modal fusion and adaptive optics.
Conclusion
DGNO constitutes a significant advance in learning-based defocus deblurring for pathological microscopy, combining the physical interpretability and robustness of discontinuous Galerkin methodologies with the adaptability of neural operators. The architecture successfully models highly spatially varying, locally discontinuous blur, producing sharper, more structurally faithful restorations that yield measurable improvement in downstream analytic tasks. This operator-theoretic paradigm demonstrates applicability beyond microscopy and paves the way for future developments in interpretable, generalizable inverse imaging with deep learning (2605.23282).
