Deep Learning of Position-Aligned Dynamic Cell Deformation Trajectories for Ovarian Cell Phenotyping in Hyperbolic Microchannels

Geometric and kinematic trajectories were extracted from three ovarian cancer cell lines (A2780, OVCAR-3, and SKOV-3) and one nonmalignant ovarian epithelial cell line (IOSE-80) at three flow rates. The trajectories were reparameterized by the axial position and resampled on a common spatial grid. Class-specific one-dimensional convolutional neural networks predicted downstream trajectories from the inlet state and flow conditions.

An unknown cell was classified by comparing its measured trajectory with class-specific predictions, and prediction errors in the width, area, perimeter, and axial velocity were integrated by soft voting. The framework achieved 93.26% accuracy, 94.07% macro precision, 90.25% macro recall, and a 91.85% macro-<i>F</i>₁ score on the held-out test data. These results support extending microfluidic cell phenotyping from isolated geometric states to the position-aligned analysis of dynamic deformation trajectories.