Robust image-based cell phenotyping via stability-guided decoupling of hydrodynamic artifacts in hyperbolic microchannels

Image-based cellular phenotyping in microfluidic channels enables high-throughput, label-free single-cell analysis. However, dynamic features extracted from cells flowing through non-uniform microchannel geometries are often influenced by operational flow conditions, which may introduce flow-associated variation into phenotypic patterns. Such flow-conditioned variation limits the robustness and generalizability of phenotyping models.Here, we present a stability-guided analytical framework for operationally decoupling hydrodynamic artifact contributions from dynamic cellular features, enabling robust image-based cell phenotyping in hyperbolic microchannels.The morphodynamic, kinematic, and cell-associated grayscale-density trajectories of normal ovarian epithelial cells and three ovarian cancer cell lines were characterized using a custom automated image-analysis pipeline, yielding a 93-dimensional dynamic feature space. To identify task-relevant descriptors, we implemented a cross-flow screening strategy based on nonparametric structural consistency and statistical stability across multiple flow rates. The analysis showed that a considerable proportion of the dynamic features exhibited substantial flow-conditioned variation.Stability-guided screening facilitated hydrodynamic artifact decoupling and reduced the original feature space to task-specific subsets containing 20 features for cancer-versus-normal classification and 25 features for cancer subtyping.Variance-attribution analysis further confirmed a substantial reduction in flow-associated effects, for example, the flow-related variance contribution to the leading principal component decreased from 69.9% to 9.3% in the subtype-classification task. Despite this dimensional reduction, the refined feature subsets maintained classification performance across multiple machine-learning models and under different data-sampling conditions.These results indicate that dynamic cellular features are highly sensitive to flow conditions, that stability-guided screening can reduce flow-conditioned artifact contributions while preserving task-relevant phenotypic information, and that stability-guided decoupling provides a robust framework for continuous-flow cellular profiling.