NIMG-65. IMPROVED SPATIAL MAPPING OF TUMOR AGGRESSIVENESS WITH 1H MAGNETIC RESONANCE SPECTROSCOPY AND DEEP LEARNING IN PATIENTS WITH NEWLY-DIAGNOSED GLIOMA

Abstract

Abstract INTRODUCTION Noninvasive, radiopathomic mapping of tumor aggressiveness can benefit patients with glioma by guiding the selection of tissue samples for diagnosis, increasing extent of resection, and non-invasively characterizing residual tumor burden for subsequent treatment. Although prior studies have demonstrated the utility of metabolic metrics quantified from 1H-MR Spectroscopy (MRS) in probing tumor pathology, this study evaluated the benefit of using the entire 1D-spectrum and deep learning for radiopathomic mapping of intratumoral cellularity, proliferation (ki-67), and a new tumor aggressiveness index (TAI) defined as log((n(ki−67)+n(cellularity))*tumor-score). METHODS Multi-voxel 1H-MRS was acquired on 281 patients newly diagnosed with a glioma (47% IDH-wildtype) immediately before surgical resection. After reconstructing individual spectra at the locations where tissue samples were obtained during surgery and normalizing by NAA in contralateral normal-appearing-white-matter, 607 spectra with corresponding histopathology were deemed of sufficient quality for analysis. A 1D convolutional-neural-network with bidirectional long- and short-term memory deep-learning model using the entire spectrum (0.6-3.6ppm) was compared to mixed-effects regression (with choline-to-NAA index[CNI]) and Random Forest (with CNI+normalized peak heights) models for predicting ki-67, cellularity, and TAI. Results & DISCUSSION Using deep-learning on the entire spectrum resulted in 10.3%-22.1% lower mean absolute error (MAE) and 0.32-0.37 higher R2 values compared to using CNI alone or a random forest model with multiple metabolic metrics. MAE values for all 3 deep-learning models were 26-44% < 1 standard deviation of the ground truth, demonstrating reasonable prediction accuracy within the test data set. Although the lowest MAE (0.16) and highest R2 (0.41) was attained when predicting TAI with deep-learning, the prediction of cellularity resulted in the lowest %MAE. Colormaps of predicted pathology identified regions of heightened aggressiveness surrounding tissue samples with most abnormal pathological features that sometimes extended beyond the non-enhancing lesion. Current work is evaluating the clinical utility of our deep-learning model and predicted maps of aggressiveness.