Incorporating the image formation process into deep learning improves network performance

Yue Li, Yijun P. Su, Min Yi Guo, Xiaofei Han, Jiamin Liu, Harshad D. Vishwasrao, Xuesong Li, Ryan Patrick Christensen et al.

Nature Methods 2022 article cited by 123

Source Open-access PDF More in Science & Research

importance 6/10 solved Advanced Image Processing Techniques microscopy imaging stage: research

Problem

Deconvolution of 3D fluorescence microscopy images is slow and artifact-prone with existing methods.

Why it matters

Faster, cleaner microscopy reconstructions accelerate biological and medical imaging research.

What the AI did

RLN, a lightweight deep network combining Richardson-Lucy iteration with a fully convolutional structure, performs fast 3D deconvolution tied to the image formation process.

Results

~16,000 parameters; 4- to 50-fold faster than data-driven networks; better deconvolution, generalizability, fewer artifacts; 4-6x faster than classic multi-view pipelines

Impact

Peer-reviewed in Nature Methods with quantitative speed and quality gains over baselines.

Record

Venue	Nature Methods
Published	2022
Work type	article
Authors	Yue Li, Yijun P. Su, Min Yi Guo, Xiaofei Han, Jiamin Liu, Harshad D. Vishwasrao, Xuesong Li, Ryan Patrick Christensen et al.
Citations	123 (OpenAlex, at import time)
References	44
Model or method	Richardson-Lucy network (RLN), fully convolutional
Organization	not stated in source
Primary source	https://doi.org/10.1038/s41592-022-01652-7
Open access	hybrid
Keywords	deep learning, deconvolution, fluorescence microscopy, Richardson-Lucy, image formation

Importance rating

6 / 10

significant — thousands to millions affected

Abstract shows large speed/quality gains across widefield, light-sheet, confocal and super-resolution imaging of cells, tissues and embryos, but states no scale of lives or fields affected.

Basis: inferred. Judged from the source text, which gives no explicit scale. Stated scale: not stated in source. Method and limits.

More in Science & Research

#	Paper	Year	Citations	Importance
1	Neural‐Symbolic Machine Learning for Retrosynthesis and Reaction Prediction Marwin Segler, Mark P. Waller	2017	549	7
2	Learning interpretable network dynamics via universal neural symbolic regression Jiao Hu, Jiaxu Cui, Bo Xiong Yang	2025	13	7
3	Data-driven Chemical Reaction Prediction and Retrosynthesis Vishnu H Nair, Philippe Schwaller, Teodoro Laino	2019	32	7
4	Generative and reinforcement learning approaches for the automated de novo design of bioactive compounds Maria V. Korshunova, Niles Huang, Stephen J. Capuzzi, Dmytro S. Radche...	2022	101	6
5	CGBVS‐DNN: Prediction of Compound‐protein Interactions Based on Deep Learning Masatoshi Hamanaka, Kei Taneishi, Hiroaki Iwata, Jun Ye, Jianguo Pei,...	2016	70	6