Related Experiment Video
Updated: Aug 12, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Improving the Neural Segmentation of Blurry Serial SEM Images by Blind Deblurring
Ao Cheng1,2, Kai Kang3, Zhanpeng Zhu4
1School of Electronic and Information Engineering, Soochow University, Suzhou 215009, China.
Computational Intelligence and Neuroscience
|January 30, 2023
Summary
This study introduces a semi-supervised learning network to deblur serial scanning electron microscopy images, improving neural connectome reconstruction accuracy. The method effectively utilizes unlabeled data, enhancing segmentation performance with fewer ground truth examples.
Area of Science:
- Neuroscience
- Computer Vision
- Machine Learning
Background:
- Serial scanning electron microscopy (sSEM) enables large-scale neural connectome reconstruction.
- Image blurring during sSEM acquisition challenges accurate neural segmentation.
- Existing deblurring methods struggle with limited training data and feature representation.
Purpose of the Study:
- To develop a novel semi-supervised learning approach for deblurring sSEM images.
- To improve the accuracy of neural connectome segmentation by enhancing image quality.
- To address limitations of current deblurring techniques in handling large datasets and high-resolution features.
Main Methods:
- Proposed a semi-supervised learning guided progressive decoding network (SGPN).
- Utilized unlabeled blurry images for training to enrich high-resolution feature representation.
- Applied deblurred datasets to train neural segmentation models.
Main Results:
- SGPN outperformed state-of-the-art deblurring models on real SEM images with reduced ground truth data.
- Achieved significant improvements in Peak Signal-to-Noise Ratio (PSNR) by 1.04 dB and Structural Similarity Index Measure (SSIM) by 0.086.
- Demonstrated substantial gains in segmentation accuracy, with A-rand decreasing by 0.119 for 2D and 0.026 for 3D segmentation.
Conclusions:
- The proposed SGPN effectively deblurs sSEM images using semi-supervised learning.
- Enhanced image quality leads to significantly improved neural segmentation accuracy.
- This method offers a more efficient approach to neural connectome reconstruction from sSEM data.
