StructRestore-3D

Self-Supervised Deep Learning for FIB-SEM Microscopy Restoration

Overview

StructRestore-3D is a research prototype for investigating self-supervised restoration of three-dimensional FIB-SEM microscopy images while preserving cellular ultrastructure.

Implemented Models

1. Single-slice convolutional neural network
2. Plain 2.5D convolutional neural network
3. Reliability-gated 2.5D convolutional neural network

Dataset

EMPIAR-10311: HeLa FIB-SEM microscopy dataset.

DOI: https://doi.org/10.6019/EMPIAR-10311

Features

• Self-supervised model training
• Masked-pixel reconstruction
• Direct target-leakage auditing
• Patch-quality diagnostics
• Multi-slice inference
• Quantitative benchmarking using PSNR and SSIM
• Neighboring-section alignment analysis

Preliminary Results

Six models were evaluated against noisy-input and conventional filtering baselines. The present models have not demonstrated superior denoising performance. Fine-structure preservation remains an important research challenge.

Requirements

Python 3.12, PyTorch, NumPy, SciPy, tifffile, scikit-image and Matplotlib.

Author

Dr. Krishna Kant Gupta

Rajiv Gandhi Institute of Information Technology and Biotechnology

Bharati Vidyapeeth (Deemed to be University), Pune, India

Research Status

Preliminary proof of concept. Further validation is ongoing.