3DUNet N2V FM2S

About

Editor:

aagamsheth

Contact email:

aagam.s6@ahduni.edu.in

Image Version:

cd1c02d6-4fca-4199-b0db-3ce33c7f9a43 — April 8, 2026

Summary

This method uses a 3-level 3D U-Net (32→64→128→128) with a large receptive field to separate noise from structure, combined with GroupNorm for stable small-batch training and residual prediction to reduce DC-shift bias. It is trained with L1 loss for sharper outputs and 1.5% uniform masking for faster N2V convergence. Mixed precision (AMP) and a cosine LR schedule improve speed and stability. The denoised output is then passed to an FM2S model for further enhancement.

Mechanism

Masked input (1.5% N2V masking) → 3D U-Net (multi-scale context, GroupNorm) → Residual prediction (input + correction) → L1 loss on masked pixels (sharp reconstruction) → AMP + cosine LR (efficient training) → FM2S model (final refinement & enhancement)

Interfaces

This algorithm implements all of the following input-output combinations:

	Inputs	Outputs
1	Stacked Synthetic Calcium Microscopy Images of Neurons, subject to noise Slug `stacked-neuron-images-with-noise` Description A .TIFF video stack of synthetic calcium imaging microscopy images depicting brain neurons. The images are subject to noise. Kind Image Read from `/input/images/stacked-neuron-images-with-noise/<uuid>.mha` or `/input/images/stacked-neuron-images-with-noise/<uuid>.tif` Stacked Synthetic Calcium Microscopy Images of Neurons, subject to noise	Stacked Synthetic Calcium Microscopy Images of Neurons, with reduced noise Slug `stacked-neuron-images-with-reduced-noise` Description A .TIFF video stack of synthetic calcium imaging microscopy images depicting brain neurons, after noise reduction. Kind Image Write to `/output/images/stacked-neuron-images-with-reduced-noise/<uuid>.mha` or `/output/images/stacked-neuron-images-with-reduced-noise/<uuid>.tif` Stacked Synthetic Calcium Microscopy Images of Neurons, with reduced noise

Validation and Performance

Left empty by the Algorithm Editors

Challenge Performance

Date	Challenge	Phase	Rank
April 8, 2026	AI4LIFE-CIDC25	Final Submission Phase: Noise Level Generalization	4
April 8, 2026	AI4LIFE-CIDC25	Final Submission Phase: Content Generalization	4

Uses and Directions

This algorithm was developed for research purposes only.

Warnings

Left empty by the Algorithm Editors

Common Error Messages

Left empty by the Algorithm Editors

Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3