Using machine-learning to optimize phase contrast in a low-cost cellphone microscope

Abstract

Cellphones equipped with high-quality cameras and powerful CPUs as well as GPUs are widespread. This opens new prospects to use such existing computational and imaging resources to perform medical diagnosis in developing countries at a very low cost. Many relevant samples, like biological cells or waterborn parasites, are almost fully transparent. As they do not exhibit absorption, but alter the light’s phase only, they are almost invisible in brightfield microscopy. Expensive equipment and procedures for microscopic contrasting or sample staining often are not available. Dedicated illumination approaches, tailored to the sample under investigation help to boost the contrast. This is achieved by a programmable illumination source, which also allows to measure the phase gradient using the differential phase contrast (DPC) [1, 2] or even the quantitative phase using the derived qDPC approach [3]. By applying machine-learning techniques, such as a convolutional neural network (CNN), it is possible to learn a relationship between samples to be examined and its optimal light source shapes, in order to increase e.g. phase contrast, from a given dataset to enable real-time applications. For the experimental setup, we developed a 3D-printed smartphone microscope for less than 100 $ using off-the-shelf components only such as a low-cost video projector. The fully automated system assures true Koehler illumination with an LCD as the condenser aperture and a reversed smartphone lens as the microscope objective. We show that the effect of a varied light source shape, using the pre-trained CNN, does not only improve the phase contrast, but also the impression of an improvement in optical resolution without adding any special optics, as demonstrated by measurements.

Towards an Intelligent Microscope: Adaptively Learned Illumination for Optimal Sample Classification

“…There are several recent works that consider use of machine learning to jointly optimize hardware and software for imaging tasks [5,6,7,8,9,10,11]. These approaches aim to find a fixed set of optical parameters that are optimal for a particular task.…”

UC2 – A Versatile and Customizable low-cost 3D-printed Optical Open-Standard for microscopic imaging

“…Using the same setup, further imaging modalities can be easily applied and tested. Due to using an LED matrix (Adafruit #1487, NY, USA) as light-source (in transmission mode) the selection of the illumination wavelength, particular pattern for contrast-maximization [45], darkfield illumination or quantitative phase-methods like "(quantitative) differential phase contrast" (qDPC, [24], see Supp. Chapter 3) and "Fourier Ptychography Microscopy" (FPM, [46]) are easily possible.…”

“…Chapter 3). For bright-field and quantitative imaging we used a 8 × 8 LED-array (Adafruit #1487, NY, USA) where a GUI was run on a 7-inch touchscreen (Raspberry Pi, UK) provides selection of individual LEDs to maximize the contrast according to Siedentopf's principle[45]. For fluorescent imaging of GFP labelled HPMEC cells, we equipped the Fluorescent-module (Supp.…”

Quantitative comparison of camera technologies for cost-effective super-resolution optical fluctuation imaging (SOFI)

“…These aim for a careful trade-off between performance and cost, instead of maximizing system performance. This goal can be achieved by a combination of open-source/open-access software and hardware blueprints, as well as repurposing commodity industrial or consumer hardware [3][4][5][6][7][8][9].…”