Biomechanics-Driven 3D Architecture Inference from Histology Using CellSqueeze3D.

Area of Science:

• Computational pathology
• Biomedical image analysis
• Quantitative tissue phenotyping

Background:

• Conventional 2D analysis of hematoxylin and eosin (H&E)-stained images is limited by tissue thickness, causing obscured cell boundaries.
• Cellular overlap and morphological changes in 2D projections hinder accurate cell size and distribution analysis.

Purpose of the Study:

• To develop CellSqueeze3D, a computational framework for reconstructing 3D cell spatial distribution and size from single H&E-stained tissue sections.
• To leverage the principle that 2D cell compression preserves 3D geometry for accurate cell reconstruction.
• To enhance computational pathology and quantitative tissue phenotyping by utilizing 3D spatial information.

Main Methods:

• CellSqueeze3D employs a hybrid Particle Swarm Optimization (PSO) approach with biomechanical constraints.
• The framework infers biologically plausible 3D cell reconstructions from 2D H&E images.
• Validation involved comparing derived nuclear-to-cytoplasmic (N/C) ratios and assessing classifier performance.

Main Results:

• The N/C ratio distribution from predicted cell radii significantly differed from random assignments (p = 1.39e-80).
• A 3D-informed cellular classifier using projected cell boundaries surpassed traditional methods, showing AUC increases of 0.136 and 0.069.
• Morphological metrics derived from CellSqueeze3D demonstrated strong associations with gene expression patterns and prognostic insights.
• Cellular and nuclear size indices predicted the mutation status of 21 genes in TCGA cohorts with a median AUROC > 0.65.

Conclusions:

• Fully utilizing 3D spatial information from single tissue slices significantly enhances computational pathology.
• CellSqueeze3D provides a novel method for accurate quantitative tissue phenotyping.
• The framework offers prognostic insights and improves prediction of genetic mutation status.