Holography, machine learning, and cancer cells

Abstract

COMMENTARY LIQUID biopsy to detect circulating solid tumor cells, cellfree DNA or tumor exosomes in peripheral blood holds great promise as a minimally invasive method to detect cancer at an early stage, plan treatments, and monitor disease response to therapy (1). Two great challenges to detecting cancer cells in whole blood are the low relative abundance of tumor cells in proportion to erythrocytes and leukocytes, and the need to introduce highly-specific exogenous labels to identify tumor biomarkers. A recently published article in Cytometry Part A (2) detects cancer cells using a rapid, label-free technique based on machine learning from quantitative phase/digital holographic microscopy datasets. The results demonstrate automated discrimination of normal and cancer cell types flowing in a microfluidic channel, pointing to the potential for high-throughput processing of circulating tumor cells in an imaging flow cytometer. The method is compatible with microfluidic approaches to sequester or enrich circulating tumor cells from normal blood cells and whole blood, respectively.Without staining, biological cells are transparent under bright field microscopy resulting in low contrast intensity images. In that article, Shaked and coauthors (2) utilize an internal contrast mechanism based on the index of refraction. Contrary to traditional qualitative phase contrast (PC) techniques, such as Zernike's PC and differential interference contrast microscopy, this off-axis interferometric label-free approach can distinguish normal cells from cancer cell lines of different disease stages using quantitative textural and morphological parameters derived from the segmented optical thickness or optical path delay (OPD) of the cell on all spatial points. This OPD is the integral of the index of refraction at each pixel of a projection across the cells' thickness. To use OPD parameters for sensitive cell screening, the phase signal must possess high fidelity and consistency across images and during the imaging session.The optical setup used in this study is a low-coherence off-axis interferometer with a Mach-Zehnder configuration. The source used is a supercontinuum laser coupled to an acousto-optical tunable filter with a 7 nm bandwidth. The low-coherence source is used to reduce speckle noise. Retroreflectors are used to adjust the sample and reference beam paths to enhance fringe visibility. Stationary aberrations and field curvature aberrations are compensated for using empty sample measurements. Phase unwrapping was performed using a standard unweighted least square algorithm to obtain the OPD maps of the cells. A normalized cut-edge detector algorithm, which is based on graph formulation, was used to separate cells from their background. This algorithm was followed by morphological image operations (opening, dilation, and thresholding) to connect the gradient lines and to remove background pixels erroneously detected. The cells from cultured cell lines, flowed through a microfluidic channel and imaged at high frame rates, ...

Quantitative assessment of cancer cell morphology and motility using telecentric digital holographic microscopy and machine learning

“…Machine learning paired with DHM has also recently been used to detect red blood cell infection by P. falciparum , discriminate between isogenic cell lines of differing metastatic stage , distinguish cancer grades from prostate biopsy microarrays , and distinguish healthy and nutrient‐deprived cancer cells , and live and dead yeast cells . Recent commentaries and reviews highlight recent studies combining these two techniques to advance cancer research .…”

On monocytes and lymphocytes biolens clustering by in flow holographic microscopy

“…Possible future applications of the presented results would be in the field of label‐free discrimination of different cell populations in cytometric approach [39–41]. There is a great demand in finding morphological biomarkers that avoid the use of fluorescent labels in order to reduce the time consumption of sample preparation and also avoid phototoxicity to allow faster and more efficient downstream analysis.…”

Deep learning of circulating tumour cells

“…So far, most machine-learning approaches for CTC analysis use traditional methods purely based on extracted features [14][15][16][17][18][19] . Only a few papers have explored deep learning methods for circulating tumour cell analysis 20,21 .…”