Nuclei instance segmentation using a transformer-based graph convolutional network and contextual information...

Area of Science:

• Medical Image Analysis
• Computational Biology
• Pathology

Background:

• Nucleus instance segmentation is crucial for biomedical applications like disease diagnosis and drug screening.
• High-density and tightly-packed cells in images present significant challenges for accurate segmentation.
• Current Convolutional Neural Network (CNN)-based methods often struggle with clustered nuclei due to limitations in bounding box regression.

Purpose of the Study:

• To develop a novel end-to-end nuclei instance segmentation model.
• To address the challenges posed by high-density and tightly-contacting cells in medical images.
• To improve the accuracy and robustness of nucleus segmentation in complex cellular environments.

Main Methods:

• Utilized Swin Transformer as the backbone for capturing global and local multi-scale features.
• Integrated a graph convolutional feature fusion module (GCFM) to learn affinity matrices and object-level local information.
• Incorporated a hybrid dilated convolution (HDC) module to enhance contextual information extraction.

Main Results:

• The proposed model demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant improvements in nucleus instance segmentation accuracy on benchmark datasets (DSB2018 and LIVECell).
• Effectively handled challenging cases of high-density and tightly-contacting nuclei.

Conclusions:

• The novel end-to-end model effectively addresses limitations in current nuclei instance segmentation techniques.
• The combination of Swin Transformer, GCFM, and HDC modules provides a robust solution for complex cellular image analysis.
• This advancement holds promise for enhanced disease diagnosis and drug screening through improved pathological analysis.