Dual Encoder Attention U-net for Nuclei Segmentation.

Area of Science:

• Computational Pathology
• Medical Image Analysis
• Deep Learning

Background:

• Nuclei segmentation in Hematoxylin and Eosin (H&E) stained whole slide images (WSIs) is crucial for computational pathology.
• Challenges include resolving touching nuclei, small nuclei, and variations in size and shape.
• Deep learning, particularly Convolutional Neural Networks (CNNs), excels at feature extraction from microscopic images.

Purpose of the Study:

• To propose a novel deep learning architecture, the Dual Encoder Attention U-net (DEAU), for enhanced nuclei segmentation.
• To improve attention mechanisms for better identification of target nuclei instances.
• To address the challenges in segmenting nuclei in H&E stained WSIs.

Main Methods:

• Developed a novel Dual Encoder Attention U-net (DEAU) architecture.
• Incorporated a pseudo hard attention gating mechanism to enhance instance attention.
• Introduced a secondary encoder using a stain-separated H channel to capture nuclei information at different resolutions.
• Evaluated the DEAU model on three public H&E datasets for nuclei segmentation.

Main Results:

• The proposed DEAU model demonstrated superior performance in nuclei segmentation tasks.
• Experimental results showed that DEAU outperforms other attention-based approaches.
• The secondary encoder effectively transforms attention priors across spatial resolutions, learning significant attention information.

Conclusions:

• The DEAU architecture offers a significant advancement in nuclei segmentation for computational pathology.
• The novel attention mechanism and dual encoder design effectively address segmentation challenges.
• This approach holds promise for automating laborious manual segmentation and counting in pathology.