Data-Centric AI Framework Improves Glioma Surgery Guidance via Fluorescence Lifetime Imaging
A novel AI framework focused on data significantly improves intraoperative fluorescence lifetime imaging (FLIm) for guiding glioma surgeries. Researchers created a strong multi-class classifier for glioblastoma (GBM) resection margins by combining confident learning (CL), class refinement, and targeted label evaluation. Data from 192 tissue margins were gathered from 31 newly diagnosed IDH-wildtype GBM patients. Initially, an expert neuropathologist categorized the data into seven tumor cellularity classes. CL assessed FLIm point-level confidence, detected labeling inconsistencies, and facilitated the merging of classes into a simplified three-class system ("low," "moderate," and "high" cellularity). This method tackles issues related to biological diversity, class imbalance, and histopathological labeling variability, aiming to enhance real-time, label-free biochemical contrast for optimal tumor resection while safeguarding functional brain tissue.
Key facts
- Data-centric AI framework integrates confident learning, class refinement, and targeted label evaluation.
- FLIm data collected from 192 tissue margins across 31 newly diagnosed IDH-wildtype GBM patients.
- Initial labeling by expert neuropathologist into seven tumor cellularity classes.
- Confident learning applied to quantify FLIm point-level confidence and identify label inconsistencies.
- Iterative class merging resulted in a three-class scheme: low, moderate, and high cellularity.
- Framework addresses biological heterogeneity, class imbalance, and histopathological labeling variability.
- Goal: improve real-time, label-free biochemical contrast for glioma surgical guidance.
- Framework aims to maximize tumor resection while preserving functional brain tissue.
Entities
—