ssEM Image Restoration via Diffusion Models With Multi-Output Joint Strategy for Noise Estimation.

Area of Science:

• Neuroscience
• Computational Biology
• Image Analysis

Background:

• Serial section electron microscopy (ssEM) is crucial for mapping neuronal connections and brain structures.
• Image degradation during ssEM acquisition presents significant challenges for accurate analysis.
• Existing deep learning methods often fail to recover high-frequency details, limiting perceptual quality and segmentation accuracy.

Purpose of the Study:

• To develop a novel diffusion model for restoring missing slices in ssEM images.
• To improve the quality and detail recovery in degraded ssEM data.
• To enhance downstream analysis tasks such as segmentation.

Main Methods:

• Utilized diffusion models for slice restoration in ssEM.
• Enhanced the backbone network with asymmetric and symmetric 3D convolutions to handle anisotropic ssEM data.
• Introduced the Adaptive and Learnable Reconstruction (ALR) module with First and Last slices Attention Block (FLAB) for feature extraction.
• Employed a Multi-output Joint Strategy (MJS) for noise estimation and diffusion correction.
• Redesigned the inference process for efficient restoration of partially damaged slices without retraining.

Main Results:

• The proposed diffusion model effectively generates more realistic ssEM slices.
• The method demonstrates superior performance in downstream tasks compared to previous approaches.
• Restoration is achieved without artifact simulation or additional retraining.

Conclusions:

• The novel diffusion model significantly improves the restoration of degraded ssEM images.
• This approach enhances the utility of ssEM data for neuroscience research by improving image quality and analytical performance.
• The method offers a robust solution for reconstructing high-fidelity brain microstructures from imperfect ssEM datasets.