OSA Fellow Bahram Jalali and his colleagues at the University of California, Los Angeles (UCLA; USA) have proposed a new time-stretch quantitative phase imaging (TS-QPI) system for capturing individual biological cells, integrated with deep machine-learning for cell classification (Sci. Rep., doi: 10.1038/srep21471). Corresponding author Claire Lifan Chen said that the technique “achieves record high-accuracy in label-free cell classification” without the potentially damaging chemical stains typically used in cell analysis. The technique could, the authors say, be used in a wide variety of applications, including data-driven cancer diagnostics, drug development, personalized genomics and biofuel production.
Time-stretch image capture
TS-QPI builds on the team’s previous work on amplified time-stretch dispersive Fourier transform (Nat. Photon., doi: 10.1038/nphoton.2012.359). With TS-QPI, an individual cell’s spatial information is encoded in the spectrum of laser pulses within a pulse-duration of sub-nanoseconds. Each pulse is then stretched in time and digitized in real time by an electronic analog-to-digital converter. The UCLA scientists report being able to capture blur-free phase and intensity images of cells at a rate of 100,000 cells per second with this new technique. (For comparison, existing flow-cytometry camera setups capture 2,000 cells per second.)
Crunching big data
TS-QPI produces massive, information-rich image datasets. The team leveraged these datasets to create an algorithm for deep machine-learning to extract 16 different features from images of individual cells—e.g., refractive index, absorption and diameter—and classify the cells in a “hyperdimensional space” composed of these 16 features. The authors say that this is the first time deep machine-learning, commonly used in speech processing and image recognition, has been used in label-free classification of cells.
Jalali and his team conducted two experimental demonstrations of TS-QPI: classifying white blood cells against colon cancer cells, and classifying lipid-accumulating algal strains for biofuel production. Their results show that compared to classification by individual biophysical parameters, their label-free TS-QPI technique improved detection accuracy from 77.8 percent to 95.5 percent.
The UCLA team worked in collaboration with scientists from NantWorks, USA, who sponsored the study.
Research News
Leveraging Big Data for Cell Imaging
Publish Date: 05 April 2016