Bayesian Transformers and Higher-Order Graph Matching for Cell Tracking in Serial Tissue Sections.

Area of Science:

• Computational Biology
• Bioimaging
• 3D Reconstruction

Background:

• 3D tissue reconstruction from serial 2D images is hindered by noise and alignment issues.
• Conventional cell tracking methods struggle with data distortions, leading to inaccurate cell linkage.
• Accurate cell tracking is crucial for understanding volumetric tissue organization.

Purpose of the Study:

• To present a novel Bayesian Transformer framework for robust cell tracking in noisy 3D tissue reconstruction.
• To improve the accuracy of linking cells across consecutive 2D sections despite image artifacts.
• To enable reliable 3D reconstruction of tissue architecture even with significant noise and distortions.

Main Methods:

• Utilized a Bayesian Transformer framework with uncertainty-aware feature embeddings.
• Incorporated higher-order graph matching with belief propagation for cell linkage.
• Employed segmentation, morphological/shape/texture descriptors, and deep CNN embeddings for feature extraction.

Main Results:

• The proposed method accurately tracks cells across serial sections, even in highly noise-prone scenarios.
• Achieved superior performance compared to baseline tracking techniques in accuracy and consistency.
• Demonstrated effectiveness on both private multiplex datasets and public time-lapse microscopy videos.

Conclusions:

• The Bayesian Transformer framework offers a robust solution for 3D tissue reconstruction from challenging multiplex imaging data.
• The uncertainty-aware approach enhances cell linkage accuracy, overcoming limitations of conventional methods.
• The method's generalizability across diverse datasets highlights its potential for broader bioimaging applications.