End-to-End Characterization of Colloidal Particles through Holographic Microscopy and Deep Convolutional Neural Networks

Analyzing holograms of colloidal particles with Lorenz-Mie
theory yields the particles' sizes, refractive indexes and three-dimensional
positions, all with exquisite precision and accuracy. No other technique
provides such a wealth of particle-resolved and time-resolved characterization
data. The underlying fits to Lorenz-Mie theory, however, require estimates
for the particles' positions and properties that are good enough to ensure
convergence to the optimal solution. Here, we demonstrate that this estimation
problem can be solved with a single, specially structured deep convolutional
neural network. The machine-learning approach to holographic particle
characterization is orders of magnitude faster than conventional image-analysis
techniques, substantially more robust against image defects, and yields answers
that already are sufficiently precise for many applications. We demonstrate
the method's efficacy through experimental measurements of the properties and dynamics
of model colloidal systems.