|
The recent progress in machine learning, a subfield of artificial intelligence (AI) with a focus on learning algorithms, is attracting researchers in quantitative phase imaging (QPI). The fast and label-free nature of QPI is ideal for generating large-scale data to train supervised machine learning algorithms. The algorithms discover important structures in large, multidimensional training data to exploit them for augmenting new QPI measurements. Here, we present two major directions in synergistically combining QPI with AI, with a particular focus on a state-of-the-art machine learning technique called deep learning.
One direction is systematic exploitation of QPI data. Employing image classification frameworks, class-dependent characteristics encoded in the images are extracted for rapid diagnosis and screening. This approach has been demonstrated in a wide range of biological systems ranging from microbes to cells to tissues, with various modalities including 2D phase, 3D tomographic, time-lapse, and spectral measurements. In these methods, AI complements limited chemical specificity of QPI by maximally utilizing refractive index information in a data-driven manner.
The second direction is an improvement of QPI methods themselves. In computational side, efficient 2D holographic or 3D tomographic reconstruction was demonstrated using neural networks. For an experimental side, reinforcement learning frameworks are employed for efficient measurements in an adaptive fashion. This direction is relatively unexplored and provides a promising frontier.
We envision that these approaches would form an indispensable toolbox for QPI and facilitate exciting new applications. As QPI is extensively studied and commercialized, rapidly accumulating data for various biological systems would render the methods increasingly powerful.
|
