SAU-Net: A Universal Deep Network for Cell Counting.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning

Background:

• Accurate cell counting is crucial for biological research but remains challenging.
• Existing methods struggle with diverse cell types and data variations.

Purpose of the Study:

• To introduce a novel deep network, SAU-Net, for universal cell counting.
• To enhance generalization in small datasets using online Batch Normalization.

Main Methods:

• Extended U-Net segmentation network with a Self-Attention module (SAU-Net).
• Developed an online Batch Normalization technique for improved data augmentation handling.
• Evaluated on four public benchmarks: VGG, MBM, ADI, and DCC datasets.

Main Results:

• SAU-Net achieved state-of-the-art performance on three real-world datasets (MBM, ADI, DCC).
• Demonstrated competitive results on the synthetic VGG dataset.
• The method shows strong generalization capabilities across different cell types and imaging conditions.

Conclusions:

• SAU-Net offers a robust and versatile solution for image-based cell counting.
• The proposed online Batch Normalization effectively addresses generalization gaps in small datasets.
• This work provides a valuable tool for advancing quantitative cell biology research.