Modeling epithelial wound closure dynamics with AI: A comparative study across cell types.

Area of Science:

• Computational biology
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• Skin wound healing is complex and heterogeneous, challenging traditional static assessments.
• Current AI methods often separate segmentation and temporal modeling, limiting dynamic quantification of healing.
• Distinct cell types exhibit unique healing trajectories that require specialized analysis.

Purpose of the Study:

• To develop an integrated AI framework for quantifying in vitro wound closure dynamics.
• To combine enhanced segmentation with temporal modeling for analyzing normal (MCF10A) and tumor (MCF7) cells.
• To compare algorithmic performance across different cell-type-specific healing phenotypes.

Main Methods:

• Implemented an enhanced UNet++ model for wound segmentation in time-lapse images.
• Modeled temporal closure trajectories using Random Forest (RF), Temporal Convolutional Network (TCN), and other regression models.
• Evaluated segmentation accuracy using Dice/IoU and temporal modeling using Mean Absolute Error (MAE) and R-squared (R²).

Main Results:

• UNet++ significantly outperformed Otsu thresholding in segmentation accuracy (p < 10^-47).
• Random Forest (RF) demonstrated superior accuracy in modeling closure trajectories for both MCF7 and MCF10A cells.
• RF effectively captured nonlinear transitions and plateau behaviors, with significant cell-type differences observed.

Conclusions:

• The integrated AI framework enables precise dynamic wound monitoring.
• This approach has clinical potential for chronic ulcer management and tumor margin surveillance.
• The framework successfully discerns cell-type-specific healing phenotypes for tailored analysis.