NIMG-83. THE PROGNOSTIC VALUE OF PARAMETRIZED KI-67 MAPS DERIVED FROM CONVENTIONAL MRI AND MACHINE LEARNING IN A LARGE RETROSPECTIVE COHORT OF ADULT-TYPE DIFFUSE GLIOMAS

Abstract BACKGROUND Ki-67 is a well-established proliferation marker relied upon for glioma grade assessment. Advances in machine learning (ML) now allow researchers to non-invasively predict the voxelwise Ki-67 index with reasonable accuracy based on anatomical MRI.The purpose of this study is to evaluate the prognostic ability of these predictions in adult-type diffuse gliomas. METHODS We included adult patients with untreated gliomas of various histopathologic grades who underwent their first resection at a comprehensive cancer center from 1993 to 2018. Anatomic MR images were pre-processed, segmented, normalized, then run through a pre-trained ML algorithm to generate Ki-67 parametric maps. The maximum estimated Ki-67 (maxKi-67) for each patient was documented, and the optimal threshold stratifying the cohort into low- and high-risk patients was computed using 10-fold cross-validation. Kaplan-Meier survival curves were then generated for both risk groups and compared using the log-rank and Breslow tests. MaxKi-67 was correlated with OS using Cox proportional hazard regression analysis adjusted for age and preoperative KPS. RESULTS 1,179 patients (age 52±15; 701 males) were included in this study. 727 tumors were histologic grade IV, while 246 and 206 were histologic grade III and II. The mean maxKi-67 was 15.9± 5.2. A maxKi-67 threshold of 13 was computed from the 10-fold cross-validation to dichotomize the cohort by OS. The low-risk group (maxKi-6713) which had median OS = 551 days (p 55 and preoperative KPS < 90 were also negative prognosticators in the Cox model: adjusted HR for age = 2.79 [2.36-3.31] and for KPS = 2.08 [1.77-2.46]. CONCLUSION Image-based derivation of maxKi-67 index using ML accurately predicts survival in glioma patients with a cut-off of 13 segregating these patients into 2 risk groups.