**4. Discussion**

In order to more specifically evaluate imaging changes consistent with treatment-related effects versus tumor recurrence, we began collecting voxel-based MRI information coupled with location specific blinded histopathological review using a within subject experimental design (i.e., contralesional matched normal voxel as a normal brain control). The goal of this project was to ultimately identify hurdles in predictive modeling regarding imaging diagnoses when longitudinally following patients with glioma after treatment to better assess true recurrence when MR changes occur, incorporating the

use of DTI into standard algorithms. Frequently, changes occur on MRI after treatment, which can be difficult to interpret. Treatments such as immunotherapy (still experimental), radiation, or cytotoxic therapy often induce changes in T2w hyperintensity and T1w contrast enhancement that can occur even several years after treatment has ended [4,50–52]. Additionally, targeted therapies, such as bevacizumab, can decrease contrast enhancement and hyperintensity, sometimes masking progression [3]. These challenges in imaging interpretation have been well known for many years. Defining progression in glioma has always been difficult and somewhat controversial.

First described by Macdonald et al., in 1990, the Macdonald criteria were imaging-based criteria to determine glioma progression based on contrast enhancement in two dimensions on CT scans in patients undergoing treatment [53]. This was later adapted to MRI and included four response categories: complete response, partial response, stable disease, or progressive disease. Macdonald criteria is limited by irregularly shaped tumors or nonspecific contrast enhancement from pharmacological treatments, radiation, inflammation, necrosis, pseudoprogression, etc. [54–56]. It also does not account for noncontrast enhancing disease, which is especially important in the evaluation of diffuse low-grade glioma. In 2010, the RANO Criteria consortium published, and later modified, guidelines for the evaluation of treatment response in gliomas and incorporated nonspecific contrast enhancement, multifocal tumors, pseudo-response after treatment, and nonenhancing fluid-attenuated inversion-recovery (FLAIR) hyperintense region in determining treatment response [57,58]. More recent measures of clinical progression have been developed to also help in distinguishing between true progression and pseudoprogression [59,60]. While these measures are important in assessing the global status of the patient and are quite sensitive and specific for global tumor recurrence, they do not answer the challenge of voxel-by-voxel analysis of imaging features specific for tumor recurrence. This study helps to further efforts of predictive, noninvasive modeling by investigating chemoradiation therapy influence on imaging in the process of determining tumor recurrence. These models can also be used to potentially better predict presence of residual disease following surgery, sites of future disease progression, and progression free survival.

This study investigated the effects of surgery alone or surgery plus radiation on voxel-specific pathology. Overall, radiation makes noninvasive differentiation of abnormal-nontumor tissue to tumor recurrence much more difficult. This is because radiation exhibits opposing behavior on key MRI modalities: specifically, on post-contrast T1, FLAIR, and GFA (a GQI feature related to FA). A number of treatment modalities clearly distinguish tumor from abnormal-nontumor postoperatively, however many of these features lose their distinguishing characteristics after radiation (see Figure 2). Specifically, features significant in both models (T1ce, FLAIR, and GFA) demonstrate contrasting information dependent on the postsurgical treatment strategy. T1ce shows that for the RT group, higher intensities indicate the presence of tumor where for the No RT group, higher intensities indicate the presence of abnormal tissue not containing tumor. The converse is true for FLAIR and GFA: for the RT group, higher intensities indicate the presence of abnormal, nontumor tissue, while lower intensities indicate tumor tissue (see Figure 3C). This implies that in order to differentiate abnormal-nontumor tissue from tumor tissue, understanding previous treatment modalities is imperative. The same approach for discriminating one for the other will not work depending on prior treatment.

Violin plots of standard MRI features (Figure 3A,B) help to understand these shifts in a more granular way. Shifts in the histograms happen all along normalized intensity values with nearly all features tested. This is predictable and influenced by treatment strategy, although histograms appear more similar after radiation, demonstrating the difficulty of distinguishing recurrence from post-treatment effects after radiation using standard features of MRI. Standard measurements also differed significantly from normal with or without radiation (Figure 2). However, distinguishing between tumor and abnormal-nontumor was difficult. FA and MD, specifically, provided no information to distinguish tumor from abnormal-nontumor tissue after radiation, although QA and GFA did. Instead, logistic regression helped to illustrate which features contributed most to differentiating the biopsy labels of tumor versus abnormal-nontumor. Hence, the opposing but important findings

described previously of T1ce, FLAIR, and GFA and their predictive value in our model. Whereas, QA (quantifies the spin orientation population in a specific direction) remained consistent across treatment models. Ultimately, both models separating images by prior treatment modality (both groups had prior surgery, some with or without chemoradiation prior to re-resection) performed well, while the aggregate model "all patients" performed poorly. This shows that the conflicting information demonstrated in Figure 3C degrades the model's ability to differentiate abnormal-nontumor from tumor tissue on MRI unless separated by treatment modality.

Overall, including non-standard DTI metrics is a useful addition towards differentiation between tumor recurrence and abnormal-nontumor MRI changes, although more is needed in the effort to improve accurate noninvasive prediction of recurrence. This study demonstrates the continued importance of matching imaging data to pathology and clinical annotation to avoid misinterpreting findings on MRI. Ultimately, combining complex datasets including pathology, genomics, epigenetics, imaging, and clinical information will all be important in improving noninvasive assessment of glioma. Future studies including more patients and more precise imaging/pathology correlation will help improve our predictive modeling to the betterment of the care of glioma patients.
