Multiscale transformers and multi-attention mechanism networks for pathological nuclei segmentation

Area of Science:

• Digital Pathology
• Medical Image Analysis
• Computer-Aided Diagnosis

Background:

• Pathology nuclei segmentation is vital for computer-aided diagnosis but challenging due to high cell density, complex backgrounds, and blurred boundaries.
• Existing methods struggle to accurately segment nuclei in complex pathological images, impacting diagnostic accuracy.

Purpose of the Study:

• To develop an advanced network model for precise pathology nuclei segmentation.
• To enhance feature extraction and boundary delineation in challenging pathological images.

Main Methods:

• A novel network model integrating a multi-scale Transformer multi-attention mechanism for pathology image segmentation.
• Incorporation of a dense attention module in the encoder to improve target cell information learning and minimize information loss.
• Inclusion of a Multi-scale Transformer Attention module between the encoder and decoder to enhance the transfer of boundary feature information.

Main Results:

• The proposed model demonstrated superior accuracy in segmenting pathology nuclei across MoNuSeg, GlaS, and CoNSeP datasets.
• The dense attention module effectively addressed challenges posed by high cell density and complex backgrounds.
• The Multi-scale Transformer Attention module significantly improved the accuracy of segmented cell boundaries.

Conclusions:

• The developed network model offers a significant advancement in pathology nuclei segmentation.
• The proposed attention mechanisms effectively overcome key challenges in segmenting dense and complex cell nuclei.
• This approach holds promise for improving the accuracy and reliability of computer-aided diagnosis in digital pathology.