Researchers develop a data-centric AI framework using fluorescence lifetime imaging to classify glioblastoma tumor margins with 96% accuracy

Fluorescence lifetime imaging (FLIm) data were collected from 192 tissue margins across 31 newly diagnosed IDH-wildtype glioblastoma patients. An expert neuropathologist initially labeled these into seven tumor cellularity classes, which were then refined into three classes ('low', 'moderate', 'high') through confident learning (a method to identify and correct label errors).

The resulting classifier achieved 96% accuracy in the three-class task. SHAP analysis (a technique for interpreting model predictions) revealed distinct optical signatures for tumor infiltration, and identified both biological factors (gray matter composition) and acquisition-related factors (blood contamination) that affected prediction confidence.

Blinded re-evaluation of margins flagged by confident learning demonstrated intra-pathologist variability, showing that selective relabeling improved data reliability more than exhaustive review of all labels.