DASHING–LSCI: deep learning-based anatomical multiclass segmentation for human intraoperative neurosurgical guidance – laser speckle contrast imaging

Abstract

Laser speckle contrast imaging (LSCI) provides real-time, label-free visualization of cerebral blood flow in vessels and perfusion in cortical tissue during neurosurgery. At present, surgeons interpret LSCI-derived flow and perfusion changes qualitatively between microsurgical maneuvers, which limits continuous quantitative analysis. Automated delineation of blood vessels and cortical tissue is a prerequisite for such analysis. Here we evaluate a deep learning approach for multiclass segmentation of intraoperative LSCI images into blood vessels, cortical tissue, and background (e.g., surgical instruments, skull). We curated 90 intraoperative LSCI images acquired during 18 unique human neurosurgery procedures across three prospective observational clinical studies (30 images per site) and trained nnU-Net (Version 2) to segment images. To assess cross-site generalization, one site was reserved exclusively for testing while the images from the other two were pooled and split into training and validation sets. Across three held-out sites, the model achieved mean Dice scores of 0.88 for background and 0.73 for cortical tissue, but a lower and more variable 0.41 for blood vessels, reflecting the difficulty of segmenting thin, sparse vessel structures. These results demonstrate the feasibility of deep-learning segmentation for extracting anatomical structure from neurosurgical LSCI images, a capability needed for real-time, continuous blood flow and perfusion analysis to support surgical guidance.