A dense multi-path decoder for tissue segmentation in histopathology images.

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning in histopathology

Background:

• Histopathological image segmentation is crucial for tissue and tumor microenvironment analysis.
• Encoder-decoder convolutional neural networks are widely used for automated segmentation.
• The role of the decoder component in these networks remains underexplored.

Purpose of the Study:

• To propose an improved decoder design for histopathology image segmentation.
• To enhance the segmentation of epithelium and stroma components.
• To investigate the impact of decoder architecture on segmentation performance.

Main Methods:

• Developed a novel decoder featuring a multi-path layout and dense shortcut connections.
• Integrated the proposed decoder with VGG, ResNet, and pre-activated ResNet encoders.
• Evaluated the method on breast and prostate histopathology image datasets (H&E stained).

Main Results:

• The network with a pre-activated ResNet encoder and the proposed decoder achieved high accuracy (ACC 0.9122, RAND 0.8398, AUC 0.9716) on breast tissue.
• Comparable high performance was observed on prostate tissue (ACC 0.9074, RAND 0.8320, AUC 0.9719).
• Achieved Dice coefficients for stroma (DICE_STR) and epithelium (DICE_EPI) above 0.90 for both datasets.

Conclusions:

• The proposed decoder design significantly improves segmentation performance compared to conventional decoders.
• The enhanced decoder contributes to more accurate automated analysis of histopathological images.
• This advancement holds potential for improving diagnostic capabilities in digital pathology.