Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FlowMorph: Physics-Consistent Self-Supervision for Label-Free Single-Cell Mechanics in Microfluidic Videos

Bora Yimenicioglu, Vishal Manikanden

TL;DR

FlowMorph tackles label-free RBC mechanical phenotyping by learning a per-track stiffness proxy $k$ through a differentiable capsule-in-flow forward model that couples laminar advection and curvature-regularized elasticity. It employs a star-convex Fourier contour representation, differentiable rasterization, and a suite of self-supervised losses (silhouette, intra-cellular flow, area conservation, wall constraints, and photometric consistency) to infer temporally coherent shapes and $k$ from short brightfield videos. A tiny calibration step maps $k$ to apparent Young's modulus $E$ via isotonic regression on a small RT-DC subset, achieving MAE $0.118$ MPa and demonstrating robustness to domain shifts. Across four public RBC datasets, FlowMorph improves physics validity, yields a strong mechanics signal distinguishing tank-treading from flipping, and provides reasonably calibrated, domain-generalizable stiffness estimates, illustrating that physics-informed forward models can enhance high-throughput, label-free single-cell mechanics analysis.

Abstract

Mechanical properties of red blood cells (RBCs) are promising biomarkers for hematologic and systemic disease, motivating microfluidic assays that probe deformability at throughputs of $10^3$--$10^6$ cells per experiment. However, existing pipelines rely on supervised segmentation or hand-crafted kymographs and rarely encode the laminar Stokes-flow physics that governs RBC shape evolution. We introduce FlowMorph, a physics-consistent self-supervised framework that learns a label-free scalar mechanics proxy $k$ for each tracked RBC from short brightfield microfluidic videos. FlowMorph models each cell by a low-dimensional parametric contour, advances boundary points through a differentiable ''capsule-in-flow'' combining laminar advection and curvature-regularized elastic relaxation, and optimizes a loss coupling silhouette overlap, intra-cellular flow agreement, area conservation, wall constraints, and temporal smoothness, using only automatically derived silhouettes and optical flow. Across four public RBC microfluidic datasets, FlowMorph achieves a mean silhouette IoU of $0.905$ on physics-rich videos with provided velocity fields and markedly improves area conservation and wall violations over purely data-driven baselines. On $\sim 1.5\times 10^5$ centered sequences, the scalar $k$ alone separates tank-treading from flipping dynamics with an AUC of $0.863$. Using only $200$ real-time deformability cytometry (RT-DC) events for calibration, a monotone map $E=g(k)$ predicts apparent Young's modulus with a mean absolute error of $0.118$\,MPa on $600$ held-out cells and degrades gracefully under shifts in channel geometry, optics, and frame rate.

FlowMorph: Physics-Consistent Self-Supervision for Label-Free Single-Cell Mechanics in Microfluidic Videos

TL;DR

FlowMorph tackles label-free RBC mechanical phenotyping by learning a per-track stiffness proxy through a differentiable capsule-in-flow forward model that couples laminar advection and curvature-regularized elasticity. It employs a star-convex Fourier contour representation, differentiable rasterization, and a suite of self-supervised losses (silhouette, intra-cellular flow, area conservation, wall constraints, and photometric consistency) to infer temporally coherent shapes and from short brightfield videos. A tiny calibration step maps to apparent Young's modulus via isotonic regression on a small RT-DC subset, achieving MAE MPa and demonstrating robustness to domain shifts. Across four public RBC datasets, FlowMorph improves physics validity, yields a strong mechanics signal distinguishing tank-treading from flipping, and provides reasonably calibrated, domain-generalizable stiffness estimates, illustrating that physics-informed forward models can enhance high-throughput, label-free single-cell mechanics analysis.

Abstract

Mechanical properties of red blood cells (RBCs) are promising biomarkers for hematologic and systemic disease, motivating microfluidic assays that probe deformability at throughputs of -- cells per experiment. However, existing pipelines rely on supervised segmentation or hand-crafted kymographs and rarely encode the laminar Stokes-flow physics that governs RBC shape evolution. We introduce FlowMorph, a physics-consistent self-supervised framework that learns a label-free scalar mechanics proxy for each tracked RBC from short brightfield microfluidic videos. FlowMorph models each cell by a low-dimensional parametric contour, advances boundary points through a differentiable ''capsule-in-flow'' combining laminar advection and curvature-regularized elastic relaxation, and optimizes a loss coupling silhouette overlap, intra-cellular flow agreement, area conservation, wall constraints, and temporal smoothness, using only automatically derived silhouettes and optical flow. Across four public RBC microfluidic datasets, FlowMorph achieves a mean silhouette IoU of on physics-rich videos with provided velocity fields and markedly improves area conservation and wall violations over purely data-driven baselines. On centered sequences, the scalar alone separates tank-treading from flipping dynamics with an AUC of . Using only real-time deformability cytometry (RT-DC) events for calibration, a monotone map predicts apparent Young's modulus with a mean absolute error of \,MPa on held-out cells and degrades gracefully under shifts in channel geometry, optics, and frame rate.
Paper Structure (35 sections, 15 equations, 2 figures, 3 tables)

This paper contains 35 sections, 15 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Overview and Notation
  5. Shape Parameterization
  6. Capsule-in-Flow Dynamics
  7. Advection by laminar flow.
  8. Curvature-regularized elastic relaxation.
  9. Wall non-penetration and no-slip.
  10. Differentiable Rasterization
  11. Temporal Inference of $s_{1:T}$ and $k$
  12. Self-Supervised Objective
  13. Silhouette consistency.
  14. Optical-flow agreement.
  15. Area conservation.
  16. ...and 20 more sections

Figures (2)

  • Figure 1: FlowMorph overview. Short brightfield clips of red blood cells (RBCs) flowing through microfluidic channels are converted into coarse silhouettes and optical flow. A lightweight encoder and temporal smoother infer a low-dimensional parametric contour and a per-track scalar mechanics proxy $k$. A differentiable physics layer advects boundary points under laminar flow plus curvature-regularized elastic relaxation, rasterizes the predicted contour into a soft mask, and computes self-supervised losses enforcing silhouette consistency, intra-cellular flow agreement, area conservation, wall constraints, and temporal smoothness. After training, a monotone map $E=g(k)$ calibrated on a small subset of real-time deformability cytometry (RT-DC) events yields approximate Young's modulus estimates.
  • Figure 2: Physics validity and mechanics signal.Top row: RBCdataset qualitative results (left) showing coarse silhouettes (green), FlowMorph contours (blue), and no-physics contours (red) overlaid on the observed flow field, and bar plots (right) summarizing IoU, area-conservation rate, wall-violation rate, and intra-cellular flow error for the methods in Table \ref{['tab:physics_rbc']}. Bottom row: TT vs. FL ROC curves (left) and box plots (right) of FlowMorph's scalar $k$ across deciles of TT posterior probability from a supervised classifier, corresponding to the metrics in Table \ref{['tab:mechanics_ttfl']}.