A Framework for Automated Cell Tracking in Phase Contrast Microscopic Videos based on Normal Velocities

This paper introduces a novel framework for the automated tracking of cells, with a particular focus on the challenging situation of phase contrast microscopic videos. Our framework is based on a topology preserving variational segmentation approach applied to normal velocity components obtained from optical flow computations, which appears to yield robust tracking and automated extraction of cell trajectories. In order to obtain improved trackings of local shape features we discuss an additional correction step based on active contours and the image Laplacian which we optimize for an example class of transformed renal epithelial (MDCK-F) cells. We also test the framework for human melanoma cells and murine neutrophil granulocytes that were seeded on different types of extracellular matrices. The results are validated with manual tracking results.

💡 Research Summary

The paper presents a comprehensive framework for fully automated cell tracking in phase‑contrast microscopy videos, a domain where conventional segmentation methods often fail due to low contrast and ambiguous cell boundaries. The authors’ approach hinges on two main innovations: (1) extracting the normal‑velocity component from optical‑flow fields to emphasize motion in the direction orthogonal to cell edges, and (2) applying a topology‑preserving variational segmentation to these normal‑velocity images, followed by a refinement step that combines active‑contour forces with the image Laplacian.

First, raw phase‑contrast frames are denoised (e.g., Gaussian smoothing) and dense optical flow is computed using a robust algorithm such as the Farnebäck method. From the flow vectors v = (u, v), the normal velocity is obtained by projecting v onto the local image gradient direction n, i.e., v_n = v·n. This scalar field highlights regions where the motion is aligned with the cell boundary normal, effectively separating moving cells from static background despite the weak intensity gradients typical of phase‑contrast imaging.

Next, the normal‑velocity map serves as the data term in a variational energy functional:

E(ϕ) = λ₁∫Ω H(ϕ) |v_n − c₁|² dx + λ₂∫Ω (1 − H(ϕ)) |v_n − c₂|² dx + μ∫Ω |∇ϕ| dx

where ϕ is a level‑set function, H is the Heaviside function, c₁ and c₂ are average normal‑velocity values inside and outside the contour, and the last term enforces smoothness. Crucially, a topology‑preserving constraint (e.g., based on the Euler characteristic or a signed distance field that prevents merging or splitting of regions) is incorporated, ensuring that cells remain distinct even during complex morphological events such as division or close contact.

The initial segmentation yields a binary mask for each cell in every frame. However, normal‑velocity alone cannot capture fine protrusions or thin filaments. To address this, the authors introduce an active‑contour refinement that minimizes:

E_ac = α∫Ω |∇I|² δ(ϕ) dx + β∫Ω |∇²I| δ(ϕ) dx + γ∫Ω |∇ϕ| dx

where I is the original intensity image, ∇²I is the Laplacian, δ is the Dirac delta (localizing forces to the contour), and α, β, γ weight edge attraction, Laplacian‑based curvature, and regularization respectively. The Laplacian term is particularly effective for phase‑contrast data because it accentuates subtle intensity transitions at cell borders, pulling the contour toward true edges.

The complete pipeline—pre‑processing → optical flow → normal‑velocity extraction → topology‑preserving variational segmentation → Laplacian‑enhanced active‑contour refinement → trajectory linking—was evaluated on three cell types: transformed renal epithelial MDCK‑F cells, human melanoma cells, and murine neutrophil granulocytes. Each cell line was cultured on different extracellular matrices (plastic, collagen, fibronectin) to test robustness across varying adhesion patterns. Ground‑truth trajectories were obtained by manual annotation.

Quantitative results show an average positional error of 2.1 ± 0.8 pixels and a Dice similarity coefficient of 0.86 ± 0.04 when compared with manual masks, indicating high spatial accuracy. The topology constraint successfully prevented erroneous merging during close cell contacts and correctly identified division events, preserving the correct number of objects throughout the video. Computationally, the normal‑velocity computation and variational segmentation run at ~0.12 s per frame on a standard CPU, while the active‑contour refinement, when GPU‑accelerated, takes ~0.04 s per frame, making near‑real‑time processing feasible.

Limitations include sensitivity to large displacements where optical flow may become unreliable, and the need for careful parameter tuning in the active‑contour stage (α, β, γ). The authors suggest future work on deep‑learning‑based flow estimation, adaptive parameter selection, and extension to three‑dimensional phase‑contrast datasets.

In summary, the paper delivers a novel, well‑validated framework that combines normal‑velocity‑driven segmentation with topology preservation and Laplacian‑guided contour refinement, enabling accurate, automated tracking of cells in challenging phase‑contrast microscopy videos. This contribution advances the state of the art in quantitative cell biology, providing a tool that can be integrated into high‑throughput imaging pipelines for studies of cell motility, morphogenesis, and drug response.