Cell Nuclei Segmentation With Dynamic Token-Based Attention Network.

Area of Science:

• Biomedical Image Analysis
• Computational Biology
• Medical Imaging

Background:

• Accurate cell nuclei segmentation is vital for disease diagnosis and treatment, enabling tasks like cell identification and classification.
• Existing convolutional neural network (CNN) methods for nuclei segmentation struggle with reliable mask prediction on diverse biomedical images.
• There is a need for advanced segmentation techniques that can handle variations in cell appearance, density, and imaging conditions.

Purpose of the Study:

• To introduce a novel Dynamic Token-based Attention Network (DTA-Net) for accurate and robust cell nuclei segmentation.
• To leverage the strengths of both CNNs and Vision Transformers (ViTs) for capturing local and global image features efficiently.
• To develop a segmentation method that minimizes computational and training costs while maximizing mask prediction reliability.

Main Methods:

• The proposed DTA-Net integrates CNNs with a ViT architecture to effectively encode both local spatial details and global contextual information.
• A novel Dynamic Token-based Attention (DTA) module is incorporated to compute attention maps, optimizing computational efficiency.
• The DTA-Net was evaluated on the 2018 Science Bowl dataset for cell nuclei segmentation.

Main Results:

• DTA-Net achieved state-of-the-art performance, demonstrating superior nuclei segmentation compared to existing methods.
• The highest Dice Similarity Coefficient (DSC) of 93.02% and Intersection over Union (IoU) of 87.91% were obtained.
• The method successfully generated high-quality segmentation masks without requiring image pre- or post-processing steps.

Conclusions:

• The DTA-Net offers a highly effective and efficient solution for automated cell nuclei segmentation in biomedical images.
• The integration of CNNs and ViTs provides a robust framework for capturing complex image features necessary for accurate segmentation.
• This approach holds significant clinical relevance by enabling reliable automated segmentation across various imaging conditions and cell types.