Enhancing cell instance segmentation in scanning electron microscopy images via a deep contour closing operator

Area of Science:

• Oncology
• Bioimaging
• Computational Biology

Background:

• Accurate cell segmentation in scanning electron microscopy (SEM) images is crucial for understanding tissue architecture in oncology.
• Current AI methods for cell segmentation in SEM images often produce errors, especially in low-quality regions, requiring extensive manual correction.
• Deficiencies in cell contour delineation hinder precise analysis of cellular structures in SEM data.

Purpose of the Study:

• To develop a novel AI-driven approach for refining cell boundary delineation in SEM images.
• To improve instance-based cell segmentation accuracy by addressing gaps in cell contours.
• To reduce the need for manual corrections in SEM image analysis for oncology.

Main Methods:

• Introduction of a convolutional neural network (CNN) Closing Operator (COp-Net) designed to fill gaps in cell contours.
• Generation of low-integrity probability maps using a partial differential equation (PDE) to overcome training data limitations.
• Validation of COp-Net using private SEM images from patient-derived xenograft (PDX) hepatoblastoma tissues and public datasets.

Main Results:

• COp-Net demonstrated significant improvements in accurately delineating cell boundaries.
• Achieved approximately 50% increase in accurately delineated cells on private data and 10% on public data compared to state-of-the-art methods.
• Substantially reduced the necessity for manual corrections, thereby accelerating the digitalization process.

Conclusions:

• The proposed COp-Net effectively enhances cell instance segmentation accuracy, particularly in challenging SEM image regions with compromised cell boundaries.
• This AI-driven gap-filling approach facilitates the detailed study of tumor tissue bioarchitecture in the field of onconanotomy.
• Public availability of COp-Net weights and PDE source code promotes reproducibility and further research.