One-click image reconstruction in single-molecule localization microscopy via deep learning

Area of Science:

• Biophysics
• Microscopy
• Computational Biology

Background:

• Deep neural networks advance microscopy image analysis, particularly in single-molecule localization super-resolution microscopy.
• Current methods require extensive manual parameter tuning and expertise, limiting model generalization and requiring retraining for new experimental conditions.

Purpose of the Study:

• To introduce AutoDS and AutoDS3D, software programs that automate single-molecule super-resolution microscopy data reconstruction.
• To significantly reduce human intervention and computational expertise needed for analyzing microscopy data.

Main Methods:

• AutoDS automatically extracts experimental parameters from raw imaging data for optimal model selection in 2D.
• AutoDS3D improves computational efficiency and integrates a graphical user interface for single-click 3D image reconstruction.
• Both methods are based on Deep-STORM and DeepSTORM3D, respectively.

Main Results:

• AutoDS removes user supervision by selecting optimal pre-trained models for 2D analysis.
• AutoDS3D offers improved computational efficiency and a streamlined workflow for 3D reconstruction.
• Both pipelines demonstrate superior performance compared to Deep-STORM and DeepSTORM3D on complex biological samples.

Conclusions:

• AutoDS and AutoDS3D significantly reduce manual labor and computation time in single-molecule super-resolution microscopy.
• These automated tools enhance the accessibility and efficiency of advanced microscopy data analysis.
• The software enables robust analysis of complex biological samples with minimal user intervention.