Modeling Wound Healing Using Vector Quantized Variational Autoencoders and Transformers

Wound healing is a fundamental mechanism for living animals. Understanding the process is crucial for numerous medical applications ranging from scarless healing to faster tissue regeneration and safer post-surgery recovery. In this work, we collect a dataset of time-lapse sequences of Drosophila embryos recovering from a laser-incised wound. We model the wound healing process as a video prediction task for which we utilize a two-stage approach with a vector quantized variational autoencoder and an autoregressive transformer. We show our trained model is able to generate realistic videos conditioned on the initial frames of the healing. We evaluate the model predictions using distortion measures and perceptual quality metrics based on segmented wound masks. Our results show that the predictions keep pixel-level error low while behaving in a realistic manner, thus suggesting the neural network is able to model the wound-closing process.