A Data-Centric Framework for Intraoperative Fluorescence Lifetime Imaging for Glioma Surgical Guidance

A team of researchers from UC Davis, led by Silvia Noble Anbunesan and including experts from the Department of Neurological Surgery, have introduced a data-centric AI framework to improve intraoperative fluorescence lifetime imaging (FLIm) for glioma surgical guidance. Published on arXiv (2604.26147), the study tackles a critical challenge in brain cancer surgery: accurately distinguishing tumor-infiltrated tissue from healthy brain in real time. The team collected FLIm data from 192 tissue margins across 31 newly diagnosed IDH-wildtype glioblastoma (GBM) patients, initially labeling them into seven tumor cellularity classes by an expert neuropathologist.

To address biological heterogeneity, class imbalance, and labeling variability, the researchers applied confident learning (CL) to quantify point-level confidence, identify label inconsistencies, and iteratively merge classes into a simplified three-class scheme ("low", "moderate", "high" tumor cellularity). This high-fidelity dataset enabled training a multi-class classifier that achieved 96% accuracy. SHAP analysis revealed distinct optical signatures across the infiltration spectrum, while targeted FLIm analysis identified biological factors (e.g., gray matter composition) and acquisition issues (e.g., blood contamination) contributing to low-confidence predictions. A blinded re-evaluation of flagged margins demonstrated intra-pathologist variability, underscoring the value of selective relabeling over exhaustive review.