Image Inpainting for Defective Microscopy Images

Recent advancements in tissue clearing and light-sheet microscopy have transformed whole-brain imaging, enabling cellular-resolution visualization of intact murine brains. However, aggressive clearing protocols and mechanical handling during sample preparation frequently introduce structural defects—such as tissue cracks and regional loss—into microscopy images. These artifacts pose significant challenges for downstream computational analyses, particularly image registration, which depends on structural continuity for precise alignment. Despite the well-documented prevalence of such defects, their impact on registration fidelity remains underexplored, and effective computational solutions for mitigating these challenges are scarce. To bridge this gap, we introduce a mask-free generative framework for digital restoration of damaged neuroimaging data. Unlike existing methods that require labor-intensive manual annotation of defect masks, our approach eliminates mask dependency entirely during both training and inference. By leveraging a diffusion-based architecture with defect-invariant learning, the model autonomously adapts to diverse defect geometries—from fine cracks to large-scale tissue loss—without prior knowledge of corruption patterns. We validate our approach using whole-brain murine microscopy datasets containing real-world artifacts induced by tissue clearing. Quantitative evaluations show that our method not only generates photorealistic restorations of missing structures but also significantly enhances registration accuracy in defective samples.