NIMG-42. MP-MRI-BASED TUMOR PROBABILITY MAPS TRAINED USING AUTOPSY TISSUE SAMPLES AS GROUND TRUTH NON-INVASIVELY PREDICT INFILTRATIVE TUMOR BEYOND THE CONTRAST ENHANCING REGION. (12th November 2021)

• Record Type: Journal Article
• Authors: Bobholz, Samuel; Lowman, Allison; Brehler, Michael; Sherman, John; Duenweg, Savannah; Kyereme, Fitzgerald; Connelly, Jennifer; Cochran, Elizabeth; Mueller, Wade; Banerjee, Anjishnu; LaViolette, Peter
• Is Part Of: Neuro-oncology. Volume 23: Supplement 6(2021)
• Page Start: vi138
• Page End: vi138
• Publication Date: 2021-11-12
• Subjects: Brain Neoplasms -- Periodicals; Brain -- Tumors -- Periodicals; Brain -- Cancer -- Periodicals; Nervous system -- Cancer -- Periodicals; 616.99481
• DOI: 10.1093/neuonc/noab196.541
• Languages: English
• ISSNs: 1522-8517

Abstract

Abstract: Infiltrative glioma beyond contrast enhancement on MRI is often difficult to identify with conventional imaging. In this study, we use large-format autopsy samples aligned to multi-parametric MRI to test the hypothesis that radio-pathomic machine learning models are able to accurately identify areas of infiltrative tumor beyond the contrast enhancing region. At autopsy, 140 tissue samples from 62 brain cancer patients were collected from brain slices sectioned to align with the patients' last clinical MRI prior to death. Cell, extra-cellular fluid (ECF), and cytoplasm densities were computed from digitized, hematoxylin and eosin-stained samples, and a subset of 20 slides from 9 patients were annotated for tumor presence by a pathologist-trained technician. In-house custom software was used to align the tissue samples to the patients' last clinical imaging, which included pre- and post-contrast T1, FLAIR, and ADC images. Bagging random forest models were then trained to predict cellularity, ECF, and cytoplasm density using 5-by-5 voxel tiles from each MRI as input. A 2/3-1/3 train-test split was used to validate model generalizability. A naïve Bayes classifier was trained to predict tumor class using cellularity, ECF, and cytoplasm segmentations within the annotation data set, again using a 2/3-1/3 train-test split to validate performance. The random forest models each accurately predicted cellularity, ECF, and cytoplasm density within the test data set, with root-mean-squared error values for each falling within one standard deviation of the ground truth. The histology-based tumor prediction model accurately predicted tumor, with a test set ROC AUC of 0.86. When using whole brain cellularity, ECF, and cytoplasm predictions from the random forest models as inputs for the naïve Bayes classifier, tumor probability maps identified regions of infiltrative tumor beyond contrast enhancement. Our results suggest that radio-pathomic maps of tumor probability accurately identify regions of infiltrative tumor beyond currently accepted MRI signatures.