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

BACKGROUND

Skin wound healing exhibits complex spatiotemporal heterogeneity that challenges traditional static assessment methods. Current artificial intelligence (AI) approaches often treat segmentation and temporal modeling as disconnected processes, limiting dynamic quantification of healing trajectories across distinct cell types. We aim to develop an integrated AI framework combining enhanced segmentation with temporal modeling for quantifying in vitro wound closure dynamics in normal epithelial (MCF10A) and tumor (MCF7) cells, and to compare algorithmic performance across healing phenotypes.

METHODS

We implemented an enhanced UNet++ model for wound segmentation in time-lapse images, benchmarked against Otsu thresholding. Temporal closure trajectories were modeled using polynomial regression, Random Forest (RF), Support Vector Regression (SVR), Autoregressive Integrated Moving Average (ARIMA), and Temporal Convolutional Network (TCN). Performance was evaluated via Dice/IoU (segmentation) and MAE/R2 (temporal modeling).

RESULTS

UNet++ achieved significantly higher segmentation accuracy than Otsu thresholding (Dice: p = 8.841 × 10-49; IoU: p = 3.931 × 10-47) with consistent temporal robustness across healing phases. For closure trajectory modeling, RF achieved superior accuracy for MCF7 (mean absolute error [MAE] = 0.48 %, R2 = 0.968) and MCF10A (MAE = 1.73 %, R2 = 0.872), excelling in capturing nonlinear phase transitions and plateau behaviors. TCN showed promise for abrupt changes in MCF7 (MAE = 1.67 %, R2 = 0.698) but failed for near-stationary MCF10A trends. Significant cell-type differences emerged, with RF providing the most interpretable predictions.

CONCLUSION

This integrated framework enables precise dynamic wound monitoring, holding clinical potential for chronic ulcer management and tumor margin surveillance, particularly through its ability to discern cell-type-specific healing phenotypes.