One-click reconstruction in single-molecule localization microscopy via experimental parameter-aware deep learning.

Area of Science:

• Microscopy
• Biophysics
• Computational Biology

Background:

• Deep neural networks (DNNs) advance microscopy image analysis, particularly in single-molecule localization microscopy (SMLM).
• DNNs predict fluorophore positions, reducing acquisition time and increasing throughput in SMLM.
• Current DNN methods require extensive manual parameter tuning for training data, limiting adaptability and increasing labor.

Purpose of the Study:

• Introduce AutoDS and AutoDS3D for automated super-resolution reconstruction in SMLM.
• Reduce human intervention in SMLM analysis by automatically extracting experimental parameters.
• Improve efficiency and user-friendliness of SMLM data processing.

Main Methods:

• Developed AutoDS and AutoDS3D based on Deep-STORM and DeepSTORM3D.
• Implemented automatic extraction of experimental parameters from raw imaging data.
• AutoDS automatically selects optimal pre-trained models for 2D SMLM; AutoDS3D enhances 3D reconstruction efficiency and integrates a GUI for one-click processing.

Main Results:

• AutoDS and AutoDS3D significantly reduce manual labor and analysis time.
• Achieved comparable or superior performance to Deep-STORM, DeepSTORM3D, and other state-of-the-art methods.
• Demonstrated complete removal of user supervision in 2D analysis with AutoDS.

Conclusions:

• AutoDS and AutoDS3D offer automated, efficient, and user-friendly solutions for SMLM super-resolution reconstruction.
• These methods overcome limitations of manual parameter tuning in DNN-based SMLM analysis.
• The software enables high-performance SMLM analysis with reduced computational cost and expert intervention.