Spatial Immunophenotyping from Whole-Slide Multiplexed Tissue Imaging Using Convolutional Neural Networks

Area of Science:

• Computational pathology
• Biomedical imaging analysis
• Deep learning applications in oncology

Background:

• Multiplexed immunofluorescence (mIF) enables detailed analysis of the tumor-immune microenvironment.
• Current mIF image analysis is labor-intensive and requires specialized pathology expertise, limiting scalability.
• There is a need for automated computational pipelines for efficient mIF analysis.

Purpose of the Study:

• To develop and validate CellGate, a deep-learning (DL) computational pipeline for automated, end-to-end whole-slide mIF image analysis.
• To streamline nuclei detection, cell segmentation, cell classification, and immuno-phenotyping.
• To enhance the scalability and clinical application of mIF analysis.

Main Methods:

• Developed CellGate, a DL pipeline for mIF image analysis.
• Trained the model on over 750,000 single cell images from 34 melanoma patients.
• Validated the pipeline on whole-slide mIF images from an independent cohort of 9 primary melanomas.

Main Results:

• CellGate demonstrated high precision-recall AUC in analyzing new whole mIF slides.
• The pipeline accurately reproduced expert pathology analysis on independent melanoma cohorts.
• Spatial immuno-phenotyping revealed insights into immune cell topography and T cell states.

Conclusions:

• CellGate provides a fully automated and parallelizable solution for whole-slide mIF image analysis.
• The DL pipeline offers improved consistency and accuracy for cell type classification.
• CellGate has the potential to enable high-throughput mIF analysis for large-scale clinical and research applications.