Which Pixel to Annotate: A Label-Efficient Nuclei Segmentation Framework.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Deep neural networks (DNNs) are prevalent in nuclei instance segmentation for H&E stained pathology images.
• Current DNNs require extensive annotated data, which is inefficient due to redundant patterns in nuclei images.
• Selective labeling strategies for nuclei segmentation remain underexplored.

Purpose of the Study:

• To propose a novel nuclei segmentation framework utilizing selective sample annotation.
• To reduce the annotation workload in deep learning for pathology image analysis.
• To achieve high-performance nuclei segmentation with minimal labeled data.

Main Methods:

• Developed a consistency-based patch selection method for identifying beneficial training samples.
• Introduced a conditional single-image Generative Adversarial Network (GAN) with a component-wise discriminator for data augmentation.
• Trained an existing segmentation model using augmented samples derived from selected patches.

Main Results:

• The proposed framework achieved performance comparable to fully-supervised methods.
• The method demonstrated effectiveness by annotating less than 5% of pixels on benchmark datasets.
• Significant reduction in annotation effort while maintaining segmentation accuracy.

Conclusions:

• Selective labeling combined with GAN-based augmentation is a viable strategy for efficient nuclei segmentation.
• The proposed framework offers a practical solution for reducing annotation costs in computational pathology.
• This approach enhances the accessibility and efficiency of deep learning in analyzing histopathological images.