BudFinder: A Masked Auto-Encoder vision transformer framework for yeast budding detection and lifespan quantification

Studying replicative aging in yeast is a central component of aging research. Recent advances in time-lapse microscopy and microfluidics now enable continuous, high-resolution tracking of individual yeast cells throughout their lifespan. However, quantifying replicative lifespan from microscopy data remains labor-intensive, as it traditionally requires manual counting of cell division events for each cell. Recent deep learning-based approaches have begun to address this challenge by automating lifespan quantification. Here, we present a versatile image analysis framework that accurately detects yeast cell division events during replicative aging. To reduce the need for large, manually annotated datasets, we pretrain a Masked Autoencoder on large-scale (~250K), unlabeled yeast cell image crops. This self-supervised pretraining substantially lowers the amount of annotated data required to train a transformer model for division event detection. Moreover, our model is trained to directly identify budding events, eliminating dependence on arbitrary heuristics such as changes in cell area. By leveraging self-supervised learning, our approach only requires training data with fewer than 50 mother cells (~1,000 division events, which is significantly lower than reported in previous methods), while maintaining high detection accuracy.