Three-dimensional localization microscopy using deep learning.

Area of Science:

• Biophysics
• Microscopy
• Computational Biology

Background:

• Single molecule localization microscopy (SMLM) is a rapidly advancing super-resolution imaging technique crucial for biosciences.
• Real-time imaging is increasingly important for observing biological dynamics, necessitating efficient data processing strategies.

Purpose of the Study:

• To introduce and evaluate a deep convolutional neural network (CNN) for 3D particle localization from single images in SMLM.
• To compare the precision and speed of the CNN approach against established methods like maximum likelihood estimation (MLE).

Main Methods:

• Development and application of a deep convolutional neural network (CNN) for 3D particle localization.
• Experimental validation using fluorescent microbeads.
• Comparison of CNN performance with maximum likelihood estimation (MLE) algorithms, including GPU-accelerated versions.

Main Results:

• The CNN achieved localization precision comparable to the gold standard MLE.
• The CNN demonstrated a speed advantage of over three orders of magnitude compared to the ThunderSTORM MLE algorithm.
• The CNN's speed becomes competitive with GPU-based MLE algorithms when estimating additional parameters, such as molecule orientation or aberrations.

Conclusions:

• CNNs offer a highly efficient and precise method for 3D particle localization in SMLM, suitable for handling large datasets.
• The CNN approach shows potential for estimating complex molecular properties and correcting for optical aberrations, further enhancing SMLM capabilities.