CellT-Net: A Composite Transformer Method for 2-D Cell Instance Segmentation.

Area of Science:

• Biomedical imaging
• Computational biology
• Artificial intelligence in healthcare

Background:

• Cell instance segmentation (CIS) is vital for cell and gene therapy, aiding in diagnosing and monitoring neurological disorders.
• Current CIS methods face challenges with cell datasets, including irregular morphology, size variations, cell adhesion, and unclear contours.

Purpose of the Study:

• To propose a novel deep learning model, CellT-Net, for effective cell instance segmentation.
• To address the limitations of existing methods in handling complex cell dataset characteristics.

Main Methods:

• Developed CellT-Net, a deep learning model utilizing the Swin transformer (Swin-T) backbone for adaptive feature focusing.
• Incorporated a novel cross-level composition (CLC) technique to enhance feature representation.
• Employed Earth Mover's Distance (EMD) and binary cross-entropy loss for precise segmentation of overlapping cells.

Main Results:

• CellT-Net demonstrated superior performance in cell instance segmentation on the LiveCELL and Sartorius datasets.
• The model effectively handled challenges like irregular cell morphology, size variation, and cell adhesion.
• Achieved precise segmentation of overlapped cells, outperforming state-of-the-art models.

Conclusions:

• CellT-Net offers an effective solution for cell instance segmentation in challenging biological image datasets.
• The proposed model advances the application of AI in cell and gene therapy for improved healthcare management.
• CellT-Net shows significant potential for clinical applications in diagnosing and tracking neurological disorders.