TA-Net: Topology-Aware Network for Gland Segmentation.

Area of Science:

• Digital pathology
• Computational image analysis
• Histopathology

Background:

• Accurate gland segmentation is crucial for quantitative analysis of gland morphology in histopathology.
• Separating densely clustered and deformed glands presents a significant challenge for current methods.
• Existing deep learning approaches using contour-based techniques have shown limited success.

Purpose of the Study:

• To propose a novel topology-aware network (TA-Net) for precise gland segmentation.
• To address the challenge of separating densely clustered and severely deformed glands.
• To enhance the generalization of gland segmentation through multitask learning.

Main Methods:

• Developed a topology-aware network (TA-Net) with a multitask learning architecture.
• Integrated gland topology estimation with instance segmentation.
• Introduced a topology loss function utilizing gland skeletons and markers to enforce topological accuracy.

Main Results:

• TA-Net achieved state-of-the-art performance on the GlaS and CRAG datasets.
• Demonstrated superior performance in segmenting densely clustered glands compared to existing methods.
• Validated using F1-score, object-level Dice coefficient, and object-level Hausdorff distance.

Conclusions:

• The proposed TA-Net effectively separates densely clustered and deformed glands.
• Multitask learning incorporating topology estimation improves gland segmentation accuracy.
• TA-Net represents a significant advancement in histopathology image analysis for gland morphology assessment.