Title: Predicting wound closure and future segmentation masks in wound healing assays
Abstract:
The prediction of the evolution of complex dynamic systems from image sequences is a significant challenge in the context of machine learning and data-driven systems, especially in the presence of non-linear dynamics, noise and limited datasets. In this work is presented a deep learning-based pipeline for automatic segmentation and temporal prediction applied to timelapse images related to wound healing assays.
The structure of the work relies in an ordered series of steps beginning with wound segmentation, followed by numerical prediction of the migration rate at predefined timepoints, and culminating with the prediction of morphological dynamics of the cellular front over time. To achieve high segmentation accuracy, a convolutional neural network based on the ResU-Net architecture has been realized to segment the wound area and to study the proliferation and migration dynamics of cells.
The numerical prediction of the cellular migration rate has been carried out using a regression model combined with a similarity procedure based on the k-Nearest Neighbour algorithm. This approach allows the system to provide accurate forecasts by referencing similar historical instances.
Furthermore, an autoregressive approach applied to the ResU-Net based convolutional neural network was used to capture the spatial evolution of cell morphology, offering a better understanding of temporal dynamics within the wound healing process.
The results demonstrate high segmentation performance and good predictive power in terms of both overall trends and morphological consistency of predictions over different time horizons. The proposed framework can be generalised to other image-based prediction problems, making it a versatile and robust tool for the study and modelling of complex dynamic phenomena in biomedical imaging.
