Cascade Semantic Segmentation by a Convolutional Neural Network in Combination with Image Super-Euclidean Pixels...

Area of Science:

• Virology
• Medical Imaging
• Computational Biology

Background:

• Accurate semantic segmentation of SARS-CoV-2 particles in electron microscopy images is underexplored.
• Existing research primarily focuses on virus detection and classification, not pixel-level delineation.
• Specialized deep learning segmentation frameworks are needed for precise viral structure identification.

Purpose of the Study:

• To develop and evaluate a deep learning system for automatic semantic segmentation of SARS-CoV-2.
• To address the gap in pixel-level delineation of viral structures using advanced segmentation techniques.
• To compare the proposed system's performance against traditional segmentation models like GVF and PIG.

Main Methods:

• A deep learning system combining convolutional neural networks (CNNs) with image processing techniques was proposed.
• The super-Euclidean pixels method was utilized as an intermediate layer within the CNN for semantic segmentation.
• Performance was benchmarked against Gradient Vector Flow (GVF) and Poisson Inverse Gradient (PIG) segmenters.

Main Results:

• The proposed CNN model demonstrated superior performance compared to GVF and PIG.
• Key metrics achieved include Dice Similarity Coefficient (DSC) of 0.9345 ± 0.0006 and Intersection over Union (IoU) of 0.8782 ± 0.0018.
• High accuracy (0.9449 ± 0.0004) and Area Under the ROC Curve (AUC) (0.9446 ± 0.0431) were reported, indicating robust segmentation capabilities.

Conclusions:

• The developed CNN-based system provides an effective automatic tool for SARS-CoV-2 segmentation in electron microscopy.
• This approach significantly surpasses traditional GVF and PIG methods in accuracy and reliability.
• Enables virologists to enhance SARS-CoV-2 detection and analysis through advanced image segmentation.