BudFinder: A Masked Auto-Encoder Vision Transformer Framework for Yeast Budding Detection.

Area of Science:

• Aging research
• Cell biology
• Image analysis

Background:

• Yeast replicative lifespan quantification is vital for aging research but is labor-intensive and time-consuming.
• Manual cell division counting is inefficient and prone to bias.
• Existing automated methods require extensive annotated data and lack adaptability across microfluidic setups.

Purpose of the Study:

• To develop a versatile and accurate image analysis approach for detecting yeast cell division events.
• To reduce the reliance on large annotated datasets for training automated division detection models.
• To improve the adaptability of automated tools for yeast lifespan studies.

Main Methods:

• Utilized a Masked Auto-Encoder pretrained on large-scale segmented yeast cell images for self-supervised learning.
• Developed a transformer model trained directly on budding event detection.
• Leveraged reduced annotated data (fewer than 50 mother cells) for model training.

Main Results:

• Achieved accurate detection of yeast cell division events.
• Significantly reduced the requirement for annotated training data (>5-fold reduction compared to prior methods).
• Maintained comparable accuracy to existing methods despite reduced data.

Conclusions:

• The proposed self-supervised image analysis approach offers a more efficient and adaptable solution for quantifying yeast cell division.
• This method lowers the barrier to entry for automated yeast lifespan analysis, particularly in microfluidic studies.
• Future applications could include broader use in aging research and cell biology studies requiring precise division event tracking.