A Fusion Model With Effective Multi-Scale Parallel Transformer for Cellular Segmentation.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• Cellular segmentation in fluorescence microscopy is hindered by uneven intensity and complex cell morphology.
• Existing models inadequately address variations in cell shape and size during segmentation.

Purpose of the Study:

• To introduce a novel network, MSPSTF-Net, for enhanced cellular segmentation.
• To integrate cell morphological information effectively into the segmentation process.

Main Methods:

• A Multi-Scale Parallel Swin Transformer (MSPST) module with 4 parallel branches captures scale-specific features.
• Multi-Scale Parallel Feature Fusion (MSPFF) and Global Feature Fusion (GFF) modules integrate morphological data.
• The proposed MSPSTF-Net was evaluated against advanced models on three biological datasets.

Main Results:

• MSPSTF-Net achieved superior segmentation performance and generalization ability.
• The model demonstrated significant improvements across F1 score, AJI, and PQ metrics.
• Our method outperformed the second-place model by an average of 1.091% (F1), 2.268% (AJI), and 1.698% (PQ) across datasets.

Conclusions:

• MSPSTF-Net offers a robust solution for challenging cellular segmentation tasks.
• The integration of multi-scale morphological features enhances segmentation accuracy and model generalization.