**3. Results**

## *3.1. Image Analysis of Biopsy Classifications*

The image data analyzed in this study is summarized in Figure 1. Eight MR/DW image features (T1w, T1ce, T2w, FLAIR FA, MD, QA, and GFA) were collected from each patient prior to re-resection. Each image feature was normalized across patients to account for fluctuations in signal acquisition due to environmental and equipment variations (Figure 1A). The image intensities were extracted from ROIs representing the locations of surgical biopsies along with their contralaterally Normal analogs (Figure 1B). Example photomicrographs of the Abnormal and Tumor biopsy classifications from one patient are displayed in Figure 1C.

**Figure 1.** A 59-year old male patient with glioblastoma multiforme. (**A**) Axial slices of the image modalities explored in this study, comprised of four standard MRI metrics (T1w, T1ce, T2w, FLAIR = fluid-attenuated inversion recovery) and four diffusion MRI metrics (fractional anisotropy (FA) and mean diffusivity (MD) quantitative anisotropy (QA) and generalized fractional anisotropy (GFA)). (**B**) Depiction of biopsies from the patient shown in (A). Filled circles indicate the locations of 0.5 mm3 Regions of Interest (ROIs) representing tissue extractions. Open circles indicate the locations of anatomically similar locations of 0.5 mm3 ROIs in the normal appearing ("healthy") contralateral hemisphere. For this patient, one biopsy (red) consisted primarily of abnormal tissue and three biopsies (magenta, cyan, and yellow) consisted primarily of tumor tissue. (**C**) Example slides of histopathology used in classification. (Left image) Tumor: Infiltrating high-grade glioma is seen with cytologically pleomorphic nuclei with large areas of necrosis and thick hyalinized blood vessels (20× magnification). (Right image) Abnormal: cortical white matter with extensive gliosis and neuropil vacuolization. Regional necrosis with thick hyalinized blood vessels consistent with radiation necrosis is present (10× magnification).

To explore the effect of radiation therapy on biopsy classification, mean signal intensities were calculated for each ROI and separated based on treatment group (Figure 2). For No RT patients (Figure 2A), differences were detected between Abnormal and Tumor in the T1ce and T2w signals (Tukey's post-hoc test, Family Wise Error Rate (FWER) = 0.05) and between Tumor and Normal in the T1w, T1ce, T2w, FLAIR, FA, and MD signals (Tukey's post-hoc test, FWER = 0.05). No differences were detected between Abnormal and Normal. For RT patients (Figure 2B), fewer image features were deemed statistically different. No differences were detected between Abnormal and Tumor (Tukey's post-hoc test, FWER = 0.05), one difference was detected between Tumor and Normal in the T1ce signal (Tukey's post-hoc test, FWER = 0.05), and two differences were detected between Abnormal and Normal in FLAIR and MD signals. The only difference consistent among treatment groups was between Tumor and Normal for the T1ce image modality; though, the feature demonstrated a reversed behavior between the two groups. Mean Normal signal intensities were equal between groups in all MRI modalities excluding MD (Figure S1).

**Figure 2.** Average ROI normalized image intensity for biopsies classified as Abnormal (green) and Tumor (blue). Contralaterally mirrored ROI locations classified as Normal (yellow). Data separated depending on chemoradiation therapy strategy prior to re-resection: (**A**) patients with adjuvant radiation therapy and (**B**) patients without adjuvant radiation therapy. Error bars show 95% confidence intervals. Asterisks indicate significance determined using Tukey's post-hoc test, *p* < 0.05.
