Few-Shot Learning for Annotation-Efficient Nucleus Instance Segmentation.

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning

Background:

• Nucleus instance segmentation in histopathology is crucial but requires laborious, expert-dependent annotation.
• Annotation-efficient deep learning methods like weakly-/semi-supervised learning are gaining interest.
• Leveraging existing fully-annotated datasets can assist segmentation on limited-annotation target datasets.

Purpose of the Study:

• To develop an annotation-efficient method for nucleus instance segmentation using few-shot learning (FSL).
• To adapt FSL for nucleus segmentation by extending it to generalized few-shot instance segmentation (GFSIS).
• To incorporate structural guidance to address challenges like touching cells and heterogeneity.

Main Methods:

• Proposed a Structurally-Guided Generalized Few-Shot Instance Segmentation (SGFSIS) framework.
• Extended few-shot instance segmentation (FSIS) to GFSIS to handle inconsistent classes between datasets.
• Integrated a structural guidance mechanism to improve segmentation accuracy for challenging nucleus features.

Main Results:

• SGFSIS demonstrated superior performance compared to other annotation-efficient methods (semi-supervised, transfer learning).
• Achieved performance comparable to fully supervised learning using only approximately 10% of annotations.
• Validated effectiveness across multiple publicly accessible histopathology datasets.

Conclusions:

• The SGFSIS framework offers a powerful solution for annotation-efficient nucleus instance segmentation.
• This approach significantly reduces the annotation burden in computational pathology.
• The method shows promise for practical applications requiring high-accuracy segmentation with minimal labeled data.