Training immunophenotyping deep learning models with the same-section ground truth cell label derivation method...

Area of Science:

• Computational pathology
• Digital pathology
• Biomarker discovery

Background:

• Deep learning (DL) models predict biomarker expression in H&E-stained tissues, aiding multi-marker immunophenotyping for cancer research and treatment.
• Current DL models often use cell labels from adjacent IHC-stained sections, which may be less accurate than labels from the same section.

Purpose of the Study:

• To evaluate the impact of cell label derivation methods on DL model performance for H&E-stained tissue analysis.
• To compare a 'same-section' labeling approach against a 'serial-section' approach using Pix2Pix generative adversarial networks (P2P-GANs).

Main Methods:

• Developed and compared two P2P-GAN virtual staining models for CD3+ T-cell prediction in lung cancer H&E images.
• One model used ground truth labels from the same tissue section, while the other used labels from an adjacent serial section.

Main Results:

• The 'same-section' DL model demonstrated significantly higher prediction performance compared to the 'serial-section' model.
• The 'same-section' model effectively stratified lung cancer patients by survival outcomes in a public cohort, indicating clinical relevance.

Conclusions:

• Ground truth cell labels derived from the same tissue section enhance the performance of DL-based immunophenotyping solutions.
• The 'same-section' approach offers a more accurate and clinically applicable method for biomarker prediction in digital pathology.