A weak-labelling and deep learning approach for in-focus object segmentation in 3D widefield microscopy

Area of Science:

• Biophysics
• Microscopy
• Computational Biology

Background:

• Biological specimens are inherently 3D, yet conventional light microscopy primarily captures 2D images.
• 3D microscopy and image reconstruction often require specialized equipment and complex techniques.
• Confocal microscopy principles inspired a novel approach to 3D widefield reconstruction.

Purpose of the Study:

• To develop a novel, accessible method for 3D widefield microscopy reconstruction.
• To improve the efficiency and speed of 3D biological image processing.
• To enable 3D imaging capabilities without specialized hardware.

Main Methods:

• Semantic segmentation of in-focus and out-of-focus pixels in widefield focal stacks.
• Application and evaluation of rule-based autofocusing algorithms for focus score mapping.
• Development of a deep neural network surrogate model for rapid pixel segmentation.

Main Results:

• Identified preferable algorithms for calculating lateral focus score maps.
• Demonstrated a computation scheme for generating focus score maps from widefield stacks.
• Achieved a significant speedup in data processing using a deep neural network model.

Conclusions:

• The proposed semantic segmentation approach enables practical 3D widefield microscopy reconstruction.
• Deep neural network integration facilitates fast and reliable online data processing.
• This method offers a more accessible route to 3D biological imaging.