Abstract

Traditional cancer diagnostics depend on time-consuming chemical staining that can delay critical surgical decisions. As part of the EU-funded CHARM project, this talk introduces a “virtual staining” pipeline that uses 38-channel broadband Stimulated Raman Scattering (SRS) microscopy to generate label-free histological images. By implementing a high-precision registration framework, we transitioned from unsupervised style transfer to a supervised deep learning approach (U-Net and CycleGAN). Our results show that while standard models provide high structural accuracy, hybrid adversarial models produce superior visual clarity and sharper cellular details. We further evaluate these virtual stains using automated nuclei segmentation (Cellpose), revealing a key divergence between human and algorithmic interpretability. This research demonstrates a chemically-grounded path for rapid, non-destructive morphological feedback in intraoperative settings, bypassing the logistical bottlenecks of conventional pathology.