PATH-60. AUTOMATED QUANTIFICATION OF KI-67 PROLIFERATION INDEX BASED ON WHOLE SLIDE IMAGES IN PEDIATRIC BRAIN TUMORS

Abstract

Abstract The use of AI methods with histopathological images (pathomics) has strong potential to aid pathologists in pediatric brain cancer diagnosis and prognosis given the morphological and molecular heterogeneity of these tumors. However, use of pathomics in this context has been relatively limited. One example is the IHC-based tumor proliferation index (percentage of Ki-67 positive cells amongst all malignant cells) which is often routinely reviewed by pathologists clinically but remains to be objectively quantified or reported in a standardized manner. As a result, analysis of this marker across patients at-scale and in data-driven methods remains largely hindered. Here, we explore the extension of an automated deep learning AI tool (DeepLIIF), originally trained and validated with adult breast/prostate/liver data, for quantification of Ki-67 index from digitized whole slide images of a cohort of pediatric brain tumors from the Children’s Brain Tumor Network (N=834). We find that derived Ki-67 scores largely agreed with neuropathologist-reported Ki-67 activity in original clinical notes and could be used to successfully stratify glioma/astrocytoma patients (N=200) into low- and high-grade groups (LGG: M=11.6; SEM=1.2; HGG: M=18.7; SEM=2; p=0.001). Ki-67 scores were also related to the survival status of the patients (Alive: M=11.8; SEM=1.1; Deceased: M=18.8; SEM=2.2; p=0.003). Summary statistics across tumor types (glioma, astrocytoma, medulloblastoma, ATRT, ependymoma, etc.) and molecularly defined subtypes are also compared, providing direct insights into distributional properties between these groups that has not previously been feasible. Our results show the ability of AI-powered methods to quantify routine histopathology metrics with little human input, allowing objective comparisons between patients and opening the door for precision medicine approaches with predictive analytics in clinical contexts.