P04.07.a AI-Based Nuclear Morphometry Reveals Distinct Morphological Patterns Between Idh-Mutant, Idh-Wildtype, and Non-Tumorous Cells in Stimulated Raman Histology

BACKGROUND Isocitrate dehydrogenase (IDH) mutation status significantly influences diffuse glioma classification. While molecular testing remains the gold standard, we investigated whether cell nuclear morphology in virtual H&E-like images of diffuse gliomas and non-tumorous tissues represents a critical feature category that could serve as a foundation for AI-based single-cell molecular subtyping using label-free Stimulated Raman Histology (SRH). MATERIAL AND METHODS Label-free and unprocessed tissue specimens were intraoperatively imaged using a portable SRH imager. We included 127 patients from a previous multicenter testing cohort, yielding 131 SRH images containing 539,145 cell regions. Specimens comprised IDH-mutant Astrocytoma (Grade 2-4), IDH-mutant Oligodendroglioma with 1p/19q-codeletion (Grade 2-3), and IDH-wildtype Glioblastoma (Grade 4). Normal tissue came from healthy brain and gliosis specimens. Cell nuclei were detected and segmented using a Mask R-CNN with a ResNet-50 backbone trained on 550 SRH-annotated patches from normal brain tissue and six different brain tumor types. We extracted three morphological features (area, perimeter, eccentricity) and analyzed their significance across cell, patch, and slide-level. RESULTS Analysis revealed significant morphological differences between IDH-mutation classes (p<0.0001). IDH-mutant cells showed larger nuclear areas (60.62±33.63μm²) and perimeters (30.21±8.92μm) versus IDH-wildtype (51.47±28.23μm², 28.27±8.48μm) and normal cells (46.67±27.69μm², 26.57±8.29μm). Eccentricity also differed significantly (IDH-mutant: 0.71±0.16, IDH-wildtype: 0.74±0.15, normal: 0.69±0.16). These differences persisted at patch and slide levels, with statistical significance by both parametric and non-parametric testing. CONCLUSION These distinctive morphological features help explain the high accuracy of current AI-based molecular SRH classifiers and demonstrate potential for single-cell molecular classification of diffuse gliomas. This capability could allow fine-granular discrimination between IDH-mutant tumor cells, IDH-wildtype tumor cells, and normal brain cells at glioma margins, potentially redefining AI-guided resection beyond current visual assessment limitations and time constraints.