SAU-Net

SAU-Net integrates a Self-Attention module into a U-Net architecture to segment and count cells in 2D and 3D microscopy images for quantitative analyses in biological research.

Key Features:

• Architecture: Extends the traditional U-Net architecture with a Self-Attention module to better capture spatial dependencies in microscopy images.
• Self-Attention: Uses self-attention to improve identification and quantification of cells within complex cellular structures and varying densities.
• Online Batch Normalization: Incorporates an online version of Batch Normalization to mitigate the generalization gap introduced by data augmentation on small datasets.
• Image support: Operates on both 2D and 3D microscopy images and is applicable to diverse cell types and densities.
• Benchmark evaluation: Evaluated on four public cell counting benchmarks: the synthetic fluorescence microscopy VGG dataset, Modified Bone Marrow (MBM) dataset, human subcutaneous adipose tissue (ADI) dataset, and Dublin Cell Counting (DCC) dataset.
• Performance: Demonstrates superior performance to state-of-the-art methods on MBM, ADI, and DCC and achieves competitive results on the synthetic VGG dataset.

Scientific Applications:

• Pathology studies: Precise cell quantification for pathological assessment and tissue analysis.
• Developmental biology: Quantification of cell numbers and spatial distributions during development.
• Tissue engineering: Evaluation of cell densities and spatial organization in engineered tissues.

Methodology:

Extends U-Net with a Self-Attention module, integrates an online variant of Batch Normalization, and evaluates performance on the synthetic fluorescence microscopy VGG, Modified Bone Marrow (MBM), human subcutaneous adipose tissue (ADI), and Dublin Cell Counting (DCC) datasets.