Efficient Metamorphosis Computation for Classifying Embryonic Cardiac Action Potentials

As purification methods for obtaining cardiomyocytes from stem cells continue to improve, the need for automated methods for highthroughput classification of these cells is becoming extremely important. Since the shape of the action potential of an adult cell is discriminative of its phenotype, a promising classification approach is to use the metamorphosis distance between the action potentials of embryonic and adult cells. However, current gradient descent methods for computing the metamorphosis distance are extremely slow, hence unsuitable for large scale classification. In this paper, we show that the metamorphosis path can be computed in closed form given the velocity field, which leads to an efficient alternating minimization approach for computing the metamorphosis distance. We test this algorithm on heart cell datasets varying from 100 to 7,000 cells.