PATH-26. Artificial-intelligence-based morphometric profiling reveals distinctive nuclear characteristics of IDH-mutant vs. IDH-wildtype tumors and non-tumorous tissue at single-cell resolution in label-free Raman imaging

Abstract BACKGROUND Isocitrate dehydrogenase (IDH) mutation status determines diffuse glioma classification, yet conventional H&E histopathological assessment cannot reliably distinguish IDH-mutant, IDH-wildtype, and non-tumor at single-cell resolution during surgery. We investigated whether artificial intelligence-based nuclear morphometry could differentiate between these three categories using label-free stimulated Raman histology. METHODS Fresh tissue specimens from 127 patients underwent intraoperative imaging using portable stimulated Raman histology, resulting in 131 images that contained 539,145 individual cell regions. Specimens included IDH-mutant astrocytomas (Grade 2-4), IDH-mutant oligodendrogliomas with 1p/19q-codeletion (Grade 2-3), IDH-wildtype glioblastomas (Grade 4), and non-tumorous tissue controls. Nuclear detection and segmentation employed a Mask R-CNN architecture with a ResNet-50 backbone, trained on 550 annotated patches representing normal brain tissue and six different tumor types. Three morphometric parameters (nuclear area, perimeter, eccentricity) underwent statistical analysis across cell and slide level. RESULTS Significant morphological nuclear differences emerged between IDH mutation categories and non-tumorous tissue (p<0.0001). IDH-mutant cells demonstrated larger nuclear areas (60.62±33.63μm²) and perimeters (30.21±8.92μm) compared to IDH-wildtype cells (51.47±28.23μm², 28.27±8.48μm) and non-tumorous cells (46.67±27.69μm², 26.57±8.29μm). Nuclear eccentricity also differed significantly: IDH-mutant (0.71±0.16), IDH-wildtype (0.74±0.15), and non-tumorous cells (0.69±0.16). These differences maintained statistical significance across all analytical levels using both parametric and non-parametric testing approaches. CONCLUSIONS Artificial intelligence-based morphometric profiling successfully identified distinctive nuclear characteristics that differentiate IDH-mutant tumor, IDH-wildtype tumor, and non-tumorous tissue at single-cell resolution. These morphological signatures can provide a foundation for molecular-guided surgery, enabling discrimination of microscopic tumor margin areas. Through refinement, this capability could support intraoperative surgical guidance by providing immediate molecular characterization at the margins of diffuse glioma, potentially optimizing the extent of resection beyond current visual assessment limitations and advancing personalized precision strategies.