Robust Machine Learning for Differentiation using 3D Point Cloud Based Features Constructed with Acceleration Vector of Embryoid Body Made from iPS Cells

Abstract

Research on induced pluripotent stem (iPS) cells is broadly categorized into two main areas: (i) replicating disease mechanisms in vitro to elucidate pathophysiology and explore new drug therapies, and (ii) efficient and safe production of iPS cells for regenerative medicine, to restore lost functions within the body. Although iPS cells have the potential to differentiate into nearly any cell type in the body, establishing standardized methods for the quantitative evaluation of differentiation remains challenging. Therefore, a method that objectively evaluates the differentiation of iPS cells is required. This paper proposes a machine learning method to assess differentiation based on the dynamic analysis of embryoid body made from iPS cells, which relies on features extracted from a three-dimensional pulsation point group derived from the acceleration and angles of embryoid body made from iPS cells pulsations over time. The process consists of the following steps: (Step 1) Utilizing Gunnar Farneback optical flow analysis to determine acceleration vectors and their directions in a hue, saturation, and value (HSV) color space from the subtle dynamics of embryoid body made from iPS cells. (Step 2) Dividing the directions of the acceleration vectors into 36 segments of 10°and summing them to form a total sum acceleration vector that represents a three-dimensional pulsation point group with respect to time and angles. (Step 3) Constructing a differentiation assessment method of embryoid body made from iPS cells using machine learning based on the objective features of periodicity, convergence, and variability extracted from the three-dimensional pulsation point group. The effectiveness of the proposed features for differentiation assessment method of embryoid body made from iPS cells was validated through computer simulations, which ascertained that the method of embryoid body made from iPS cells achieved an accuracy of 84%and F-score of 80%. This highlights the robustness and high precision of the proposed machine learning-based differentiation assessment method of embryoid body made from iPS cells, which focuses on the quantitative features of the acceleration vectors within three-dimensional pulsation point groups.