Efficient Staining-Invariant Nuclei Segmentation Approach Using Self-Supervised Deep Contrastive Network.

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Nuclei segmentation in hematoxylin and eosin (H&E) whole slide imaging (WSI) is crucial but challenging due to staining variations and diverse nuclei morphology.
• Existing methods often rely on stain normalization, which can lead to information loss and fail to address inter-scanner variability.

Purpose of the Study:

• To develop an efficient and robust nuclei segmentation method for H&E WSI that is invariant to staining variations.
• To overcome the limitations of stain normalization and inter-scanner feature instability in current segmentation approaches.

Main Methods:

• A staining-invariant encoder (SIE) incorporating convolution and transformer blocks was developed.
• A weighted hybrid dilated convolution (WHDC) block was introduced to capture multi-scale nuclei features, accommodating variations in size and shape.
• The SIE was trained using self-supervised contrastive learning on five unlabeled WSI datasets, serving as a backbone for the segmentation network.

Main Results:

• The proposed method demonstrated superior performance in nuclei segmentation across multiple challenging WSI datasets.
• The approach successfully segmented nuclei without requiring a stain normalization step.
• The WHDC block effectively handled variations in nuclei size and shape, improving segmentation accuracy.

Conclusions:

• The developed staining-invariant nuclei segmentation method offers a robust solution for H&E WSI analysis.
• Self-supervised learning combined with the WHDC block provides an effective strategy for nuclei segmentation, eliminating the need for stain normalization.
• This method enhances the reliability and applicability of automated nuclei segmentation in digital pathology.