VirtualPNN: Deep Learning for Label-free Visualization of Perineuronal Nets in Brain Tissue
Year: 2025
Type: article
Abstract: Perineuronal nets (PNNs) are extracellular matrix structures that modulate inhibitory interneuron activity and may contribute to brain tumor–related epilepsy. Because immunohistochemistry (IHC) destroys tissue and is time-intensive, this study introduces VirtualPNN, a generative AI model that simulates aggrecan IHC staining directly from H&E slides. Trained using paired H&E–IHC images and an anti-vessel loss objective, VirtualPNN produced high-fidelity stains, outperforming AI-FFPE, CycleGAN, CUT, and FastCUT, as demonstrated by the lowest Fréchet Inception Distance score. A visual Turing test with board-certified neuropathologists showed high fooling rates, confirming realistic outputs. Importantly, the model minimized false-positive staining of blood vessels through its anti-vessel loss component. VirtualPNN offers a label-free, tissue-preserving method to study PNNs and may serve as a tool for correlating PNN distribution with epilepsy severity in tumor patients. Ongoing work includes validation across larger cohorts and extension to additional staining markers.
Institutions: Dartmouth College, Dartmouth–Hitchcock Medical Center
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203904718