Automatic segmentation for synchrotron-based imaging of porous bread dough using deep learning approach.

Area of Science:

• Materials Science
• Biotechnology
• Data Science

Background:

• Advancements in synchrotron beamlines enable high-resolution, rapid 3D imaging of dynamic processes.
• Analyzing large 3D datasets from experiments like bread dough rising presents significant segmentation challenges.
• Current semi-automated segmentation methods are labor-intensive and time-consuming.

Purpose of the Study:

• To explore machine learning for automating 3D image segmentation in synchrotron tomography.
• To address the challenge of generating adequate training data for machine learning models.
• To improve the efficiency and accuracy of analyzing dynamic structural changes.

Main Methods:

• Development of methods for automatically generating synthetic training datasets with realistic image attributes.
• Training a U-Net machine learning model using the generated synthetic data.
• Application of the trained U-Net model for segmenting 3D tomographic images of bread dough during rising and baking.

Main Results:

• The U-Net model trained on synthetic data successfully segmented complex 3D bread dough structures.
• The automated machine learning approach significantly outperformed previous semi-automated segmentation techniques.
• The U-Net model required substantially less manual user input compared to existing methods.

Conclusions:

• Automated 3D image segmentation using machine learning, particularly U-Net trained on synthetic data, is effective for synchrotron tomography.
• This approach reduces manual effort and accelerates the analysis of large datasets.
• The developed method facilitates 4D characterization studies with finer temporal resolution.