Multiresolution semantic segmentation of biological structures in digital histopathology.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Semantic segmentation of histopathology whole slide images (WSIs) is crucial for pathology applications.
• Patch-based Convolutional Neural Networks (CNNs) are state-of-the-art but face challenges with varying structure sizes and resolution trade-offs.
• Existing methods struggle to balance field-of-view, computational efficiency, and spatial resolution.

Purpose of the Study:

• To propose a multiresolution semantic segmentation approach for histopathology WSIs.
• To address the trade-offs between field-of-view, computational efficiency, and spatial resolution.
• To improve the accuracy and consistency of segmenting biological structures of diverse sizes.

Main Methods:

• A two-stage multiresolution approach using CNNs was developed.
• The first stage employs four CNNs for feature extraction at different resolutions.
• The second stage uses a CNN to aggregate features and generate segmentation masks.

Main Results:

• Achieved 95.6% pixel-wise accuracy on a placenta dataset and 97.1% on a lung dataset.
• Reported a mean Dice similarity coefficient of 92.5% (placenta) and 87.3% (lung).
• Obtained a mean positive predictive value of 97.1% (placenta) and 83.3% (lung).

Conclusions:

• The multiresolution approach demonstrates high accuracy and consistency in segmenting varied biological structures.
• The method is effective for both single-class (placenta) and multiclass (lung) histopathology WSIs.
• This technique can advance automated analysis in histopathology, aiding clinical research.