Abstract

In tumor resection surgery, the goal is complete removal of malignant tissue while conserving the surrounding normal tissue. As gross visual inspection does not offer adequate sensitivity for demarcating the tumor boundary, the success of these procedures is confirmed by microscopic analysis using gold standard H&E histopathology methods. This analysis is currently performed post-operatively given the time-intensive processing required to prepare thin H&E-stained tissue sections from an excised specimen as outlined in Fig. 1a, where positive margins may necessitate an additional re-excision surgery. Alternatively, frozen section analysis can be used to obtain intraoperative feedback for surgeons, though freezing artifacts can lead to diagnostic errors, and tumors in lipid-rich or calcified tissues are typically poor candidates for freezing. Microscopic imaging for rapid, accurate intraoperative histology remains an unmet need, with the potential to significantly improve re-excision rates. Our imaging approach known as ultraviolet photoacoustic remote sensing (UV-PARS) microscopy is sensitive to DNA optical absorption at 266 nm in a non-contact, reflection-mode configuration.^1,2 This provides label-free contrast to cell nuclei, comparable to a hematoxylin stain. This method is easily integrated with 266 nm optical scattering microscopy with contrast resembling an eosin stain.³ Images in unstained tissues using this dual-channel system can be pseudo-colored and combined for a complete virtual histological visualization, with features that are concordant with conventional H&E staining. However, a disparity in stain style remains between virtual histology and the gold standard. In this work, we leverage deep learning-based image style transfer methods using a cycle-consistent generative adversarial network (CycleGAN), allowing the input pseudo-colored virtual histology to be rendered in a maximally realistic stain style familiar to pathologists. This approach aims to aid diagnostic interpretation, facilitate integration into existing clinical workflows, and ensure compatibility of our virtual histological images with the breadth of existing diagnostic AI methods developed for conventional H&E-stained histopathology.

© 2023 SPIE