Applications of deep learning to the assessment of red blood cell deformability

Abstract

BACKGROUND: Measurement of abnormal Red Blood Cell (RBC) deformability is a main indicator of Sickle Cell Anemia (SCA) and requires standardized quantification methods. Ektacytometry is commonly used to estimate the fraction of Sickled Cells (SCs) by measuring the deformability of RBCs from laser diffraction patterns under varying shear stress. In addition to estimations from model comparisons, use of maximum Elongation Index differences (ΔEImax) at different laser intensity levels was recently proposed for the estimation of SC fractions. OBJECTIVE: Implement a convolutional neural network to accurately estimate rigid-cell fraction and RBC concentration from laser diffraction patterns without using a theoretical model and eliminating the ektacytometer dependency for deformability measurements. METHODS: RBCs were collected from control patients. Rigid-cell fraction experiments were performed using varying concentrations of glutaraldehyde. Serial dilutions were used for varying the concentration of RBC. A convolutional neural network was constructed using Python and TensorFlow. RESULTS: Our measurements and model predictions show that a linear relationship between ΔEImax and rigid-cell fraction exists only for rigid-cell fractions less than 0.2. Our proposed neural network architecture can be used successfully for both RBC concentration and rigid-cell fraction estimations without a need for a theoretical model.

Critical review of single-cell mechanotyping approaches for biomedical applications

“…In doing so we highlight developing approaches that combine low fabrication cost, microfluidic cell handling and the ability to modulate deforming forces, 52,75 in addition to recent developments supporting the potential for machine learning to enhance characterization. [76][77][78][79][80] We also identify directions for future work to improve the throughput, data quality and reliability of single-cell mechanotyping measurements.…”

“…Machine learning techniques for single-cell mechanotyping. The potential for machine learning techniques to advance single-cell mechanotyping technology has recently been proposed, namely as a complimentary data analysis tool utilizing any number of cell deformation techniques, [76][77][78][79]190,191 or as a standalone system capable of estimating cell deformability via associated morphological features, free from any form of mechanical probing. 80 Such machine learning approaches offer a methodology for characterising intrinsic cell stiffness properties, which is otherwise complicated owing to the complex biophysical nature and interplay between cellular components, as well as difficulties in calibration and standardization of mechanotyping load cases.…”

“…Machine learning tools are thus useful in characterising mechanical phenotypes where the relationship between cell stiffness and raw experimental data is nonlinear, noisy or otherwise complex. Machine learning algorithms applied in this manner have accordingly been trained to classify mechanotypes using complex cell trajectories under deformability-induced lift forces, 76 laser diffraction patterns under fluid shear forces, 79 optical images from RT-DC analysis 190 and acoustofluidic deformation forces. 77,191 The potential for machine learning techniques to identify mechanical characteristics from optical analysis alone has also been investigated.…”