Abstract A066: Scalable visualization of perineuronal nets in the brain tumor microenvironment using generative deep learning

Key Points

• The research aims to develop a scalable method for visualizing perineuronal nets in brain tumors to understand their role in epilepsy better.
• Developed VirtualPNN, a generative deep learning framework for imputing PNN staining from histopathology samples.
• Analyzed a cohort of 133 tissue microarray cores from 44 patients with various brain tumors.
• Evaluated model performance using quantitative similarity metrics and blinded neuropathologist review.
• Incorporated domain-informed constraints to reduce false positives from vascular signals.
• Conducted robustness testing in tumor tissues lacking PNN morphology.
• VirtualPNN showed the lowest Fréchet Inception Distance compared to other AI models, indicating strong similarity to true immunohistochemical staining.
• The model preserved identifiable PNN morphology and reduced off-target vascular staining significantly.
• Patient-level analysis demonstrated continuous feature-space concordance with real PNN staining across seizure history.
• Robustness tests revealed lower spurious signal generation than baseline models, particularly in hypocellular regions.