Deep learning-based size prediction for optical trapped nanoparticles and extracellular vesicles from limited...

Area of Science:

• Biophysics
• Nanotechnology
• Optical Physics

Background:

• Camera-based monitoring in optical traps offers multi-parametric characterization of nanoparticles.
• Inaccurate size predictions arise from blurring and limited temporal bandwidth in current methods.
• Strong optical trap stiffness exacerbates position detection errors.

Purpose of the Study:

• To develop an accurate size characterization method for trapped nanoparticles.
• To overcome limitations of existing sizing algorithms in optical trapping.
• To enable precise morpho-optical characterization at the single-particle level.

Main Methods:

• A ResNet-based deep learning model was trained using simulated time series data of constrained Brownian motion.
• The method utilizes camera-based monitoring of nanoparticles within optical traps.
• Combined sizing network with high-speed video still frames for size and refractive index quantification.

Main Results:

• The ResNet method outperforms state-of-the-art sizing algorithms like adjusted Lorentzian fitting and CNNs.
• Accurate sizing was achieved for standard nanoparticles and extracellular vesicles (EVs), even with short measurement times (<1s).
• The network accurately determined EV size distribution in clinical samples, validated against nanoparticle tracking analysis (NTA).

Conclusions:

• The proposed ResNet-based method enables accurate size characterization of trapped nanoparticles and EVs.
• This approach overcomes blurring and aliasing issues inherent in camera-based optical trapping.
• The technique offers a robust path for predicting bio-nanoparticle morphological heterogeneity and quantifying size and refractive index.