Research map / 3D vision / Depth and stereo
A Gaussian Filter-Based 3D Registration Method for Series Section Electron Microscopy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a Gaussian-filter-based 3D registration method for ssEM that extends a 1D Gaussian filter to image stacks and combines it with a cascaded pyramid optical flow network to remove nonlinear distortions while preserving natural deformations.
Problem
3D registration of series section electron microscopy stacks must remove axial nonlinear distortions from sectioning while preserving natural biological deformations, and avoid error accumulation over long sequences.
Approach
Reformulates 3D registration as a superposition of high- and low-frequency components and applies a 1D Gaussian filter over deformation fields between neighboring slices (integrated with a cascaded pyramid optical flow network), with a multi-neighbor weighted similarity loss.
Key results
- On Mouse Kidney, the method reaches MI 1.54, SSIM 0.78, NCC 0.99 vs. best baseline EFSR at 1.26 / 0.63 / 0.97 (GT comparison).
- Average per-slice time is 0.72 s versus 3.52 s for TrakEM2, 3.38 s for EFSR, 0.25 s for EMReg, and 29.04 s for SEMLeSS.
- Gaussian filter ablation: with filter 0.78 vs. 0.63 without on Mouse Kidney; LKUnet improves from 0.43 to 0.75 when the filter is added.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Gaussian filter strategy (ablation) | Ours · Mouse Kidney | Registration accuracy: 0.63 -> 0.78 | Helped |
| add the Gaussian filter strategy (ablation) | Ours · Mouse Skin | Registration accuracy: 0.67 -> 0.82 | Helped |
| add the Gaussian filter strategy (ablation) | Ours · Mouse Liver | Registration accuracy: 0.66 -> 0.80 | Helped |
| replace the optical flow estimation network with LKUnet while retaining Gaussian filtering (ablation) | Ours · Mouse Kidney | Registration accuracy: 0.78 -> 0.75 | Hurt |
| replace the optical flow estimation network with LKUnet while retaining Gaussian filtering (ablation) | Ours · Mouse Skin | Registration accuracy: 0.82 -> 0.80 | Hurt |
| replace the optical flow estimation network with LKUnet while retaining Gaussian filtering (ablation) | Ours · Mouse Liver | Registration accuracy: 0.80 -> 0.77 | Hurt |
| increase Gaussian kernel radius from r=1 to r=2 (ablation) | Ours · Mouse Kidney | Registration accuracy: 0.83 -> 0.72; 0.82 -> 0.65; 0.83 -> 0.92 | Mixed |
| increase Gaussian kernel radius from r=1 to r=3 (ablation) | Ours · Mouse Kidney | Registration accuracy: 0.83 -> 0.58; 0.82 -> 0.55; 0.83 -> 0.58 | Hurt |
| vary iteration count T from 1 to 3 at r=1 (ablation) | Ours · Mouse Kidney | Registration accuracy: 0.83; 0.82; 0.83 | No clear effect |
| use the proposed method instead of TrakEM2 | 3D registration · Mouse Kidney | SSIM (GT): 0.61 ± 0.03 -> 0.78 ± 0.03 | Helped |
| use the proposed method instead of TrakEM2 | 3D registration · Mouse Liver | SSIM (GT): 0.61 ± 0.03 -> 0.80 ± 0.02 | Helped |
| use the proposed method instead of TrakEM2 | 3D registration · Mouse Skin | SSIM (GT): 0.64 ± 0.06 -> 0.82 ± 0.03 | Helped |
Surprising findings
- Increasing the Gaussian receptive field radius r from 1 to 2 or 3 sharply reduces registration accuracy, and EMReg is faster than the proposed method despite lower accuracy.
Limitations
- Fixed Gaussian kernel with radius 1 is used; larger receptive fields degrade accuracy.
- Evaluation relies on synthetic elastic distortions applied to real EM data with ground truth.
Experimental setup
Datasets: OpenOrganelle mouse kidney OpenOrganelle mouse liver OpenOrganelle mouse skin
Benchmarks: OpenOrganelle mouse kidney OpenOrganelle mouse liver OpenOrganelle mouse skin
Metrics: NCC SSIM MI time per slice
Baselines: TrakEM2 EMReg EFSR SEMLeSS LKUnet
Builds on: optical flow estimation (RAFT-style recurrent update) EMReg Gaussian filtering
Closest papers by meaning
- V2V3D: View-to-View Denoised 3D Reconstruction for Light Field Microscopy (CVPR 2025 · similarity 0.84)
- At FullTilt: Real-Time Open-Set 3D Macromolecule Detection Directly from Tilted 2D Projections (NeurIPS 2026 · similarity 0.84)
- BOE-ViT: Boosting Orientation Estimation with Equivariance in Self-Supervised 3D Subtomogram Alignment (CVPR 2025 · similarity 0.84)
- Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery (NeurIPS 2025 · similarity 0.84)
- Volume Tells: Dual Cycle-Consistent Diffusion for 3D Fluorescence Microscopy De-noising and Super-Resolution (CVPR 2025 · similarity 0.83)
- Orientation-anchored Hyper-Gaussian for 4D Reconstruction from Casual Videos (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.