Cell2Grid: an efficient, spatial, and convolutional neural network-ready representation of cell segmentation data.

Area of Science:

• Computational pathology
• Digital image analysis
• Machine learning in histopathology

Background:

• Cell segmentation algorithms aid histologic image interpretation but hinder spatial analysis.
• High-resolution imaging for convolutional neural networks (CNNs) is computationally intensive.
• A need exists for efficient spatial data representations for CNNs.

Purpose of the Study:

• To propose and investigate Cell2Grid, an alternative spatial data representation for direct CNN training.
• To enable hypothesis-free spatial analysis using cell segmentation data.
• To overcome computational limitations of high-resolution cell-based imaging.

Main Methods:

• Developed Cell2Grid algorithm to generate compact images from cell segmentation data.
• Placed individual cells into a low-resolution grid, resolving cell conflicts.
• Evaluated Cell2Grid using a colorectal cancer relapse prediction case study on multiplex immunohistochemistry images.

Main Results:

• Cell2Grid images were 100x smaller than original segmentation data.
• Key cell features (phenotype counts, nearest-neighbor distances) were preserved.
• CNNs trained on Cell2Grid images reduced test set error rate by 25% and training time by 85% compared to other methods.

Conclusions:

• Cell2Grid is an efficient algorithm for using CNNs with cell segmentation data.
• The cell-based representation facilitates model interpretation and synthetic image generation.
• Cell2Grid enhances spatial analysis in computational pathology.