Importance of dataset design in developing robust U-Net models for label-free cell morphology evaluation

Area of Science:

• Biomedical Engineering
• Computational Biology
• Regenerative Medicine

Background:

• Label-free cell image analysis is crucial for non-invasive cell quality evaluation in regenerative medicine.
• Automated, image-based cell analysis offers efficiency and quantitative insights, but segmentation remains a challenge.
• Limited data availability in cell culture poses difficulties for developing robust deep learning models.

Purpose of the Study:

• To investigate how training dataset design impacts the robustness of U-Net models for cell segmentation.
• To determine the optimal dataset characteristics for effective cell segmentation with limited data.
• To provide guidance for deep learning model development in cell manufacturing and research.

Main Methods:

• Trained U-Net models using 2592 image pairs from four cell types.
• Evaluated 42 dataset design patterns focusing on size, content, and morphological diversity.
• Assessed model performance based on segmentation accuracy and robustness across different cell types.

Main Results:

• Robust cell segmentation models can be achieved with as few as 10 raw images (4× objective).
• Training datasets featuring common cell patterns yield more robust models than those with rare patterns.
• Including diverse morphologies (spindle and round cells) significantly enhances model robustness across cell types.

Conclusions:

• Dataset content and morphological diversity are critical factors for robust deep learning-based cell segmentation.
• Careful curation of training datasets is essential for effective cell image analysis.
• This study offers practical insights for optimizing dataset design in cell manufacturing and research.