Taming Glioblastoma in “Real Time”: Integrating Multimodal Advanced Neuroimaging/AI Tools Towards Creating a Robust and Therapy Agnostic Model for Response Assessment in Neuro-Oncology

Glioblastoma is the most common primary malignant brain tumor in adults and carries a dismal prognosis. Despite maximal safe surgical resection followed by concurrent chemoradiotherapy (CCRT) using temozolomide and adjuvant temozolomide, the prognosis remains very poor, with median overall survival (OS) of 15 to 17 months (1). To improve outcomes, early and accurate diagnosis of postoperative tumor progression (TP) is crucial because it affects clinical decisions for therapeutic “early” intervention and determining overall prognosis. Pseudoprogression can confound therapeutic decision-making as these patients often have methylated O⁶-methylguanine-DNA-methyltransferase (MGMT) promoter, respond well to temozolomide, and have superior OS compared with patients with TP (2). These patients are closely monitored with short-interval follow-up MRI scans and continue on adjuvant temozolomide. Conversely, patients with TP often require repeat biopsy/surgical resection and/or switching to alternative therapies such as tumor treating fields (TTFields), chemowafers, antiangiogenic therapy, or immunotherapy (3).

Pseudoprogression (PsP), aka “treatment-related change”, is most commonly seen within 12 weeks after the completion of CCRT, characterized by geographic necrosis, reactive gliosis, and vascular hyalinization, but it may occur later in the course of the disease, up to six months after treatment (4). Both TP and PsP can exhibit new/increasing enhancement within the radiation field with progressive enlargement of T2/FLAIR signal abnormality. Importantly, these morphologic changes reflect impairment of the blood-brain barrier and are, therefore, nonspecific surrogate markers, and can be seen in both TP and PsP. To this end, revised criteria for assessment of treatment response in high-grade gliomas were proposed by the RANO working group (5, 6) to distinguish PsP from TP.

The modified RANO (mRANO; ref. 5) criteria, as the former RANO criteria (6), include measurement of enhancing lesions via 2D-biperpendicular diameter and use of percentage thresholds to define response and progression, both depending on serial follow-up MRI examinations. The updates of mRANO include the removal of qualitative nonenhancing tumor assessment requirements by T2-FLAIR abnormality and the use of the postradiation time point as the baseline for newly diagnosed glioblastoma response assessment. RANO/mRANO criteria have remarkably impacted the retrospective examination in clinical trials (7) and could also be adapted for routine clinical practice. Nevertheless, advanced MRI techniques, including diffusion tensor imaging (DTI), dynamic susceptibility contrast-perfusion weighted imaging (DSC-PWI), dynamic contrast enhanced (DCE)-MRI and magnetic resonance spectroscopy (MRS), widely used as a routine in clinical neuro-oncologic imaging in many major centers, can substantively support conventional MRI findings not only in “real-time” but also in a more objective quantitative manner.

According to some studies (8, 9) there is a reasonable agreement between bidimensional (2D) and volumetric (3D) measurements of contrast-enhancing tumors, advocating for 2D measurements as a more practical alternative. However, higher reader discordance using bidirectional measurements has also been quoted by several studies (10–12), particularly in high-grade gliomas (5). This fact could reflect the current challenges of standard MRI interpretation to monitor glioma growth, (i) most enhancing glioblastomas have an irregular morphology and/or ill-defined margins and can also present with multifocal lesions (13); (ii) PsP affects vascular permeability, increasing the area of enhancement within the radiation field, overestimating tumor volume (13); (iii) antiangiogenic agents, such as bevacizumab, promote a marked decrease in contrast enhancement but no significant difference in OS, known as pseudo response, overestimating response rate on the basis of the use of contrast enhancement as the principal instrument (14, 15). Collectively, the increased regional heterogeneity observed on follow-up imaging makes 2D as well as 3D measurements less specific, thereby warranting the exploitation of additional advanced neuroimaging techniques and artificial intelligence (AI) tools in a multiparametric approach.

Malignant gliomas exhibit heterogeneous biological phenotypes, with areas of both neovascularized (enhancing) and infiltrative (nonenhancing) components (16). Due to the subjective nature of the interpretation of T2-FLAIR, this sequence was removed from the mRANO criteria to examine nonenhancing tumor growth. However, it is well known that nonenhancing tumor progression results in dramatically different postprogression survival in patients with glioblastoma (17); nevertheless, T2-FLAIR has a limited role in differentiating inflammatory changes versus tumor infiltration, confounding the response assessment, thereby warranting association with advanced MRI metrics, such as DSC-PWI, DTI, and MRS.

The use of cutoffs is necessary to standardize measurement comparison among different centers, and RANO/mRANO criteria rely on thresholds to define response and progression (≥50% decrease for partial response; PR and > 25% increase for progressive disease; PD in the sum of product diameters). Ellingson and colleagues (5) reported relative arbitrariness of RANO thresholds, which are not optimized on the basis of scientific data to show the best correlation with survival benefit or time to treatment failure. Also, using thresholds based on percentage change is particularly biased for small tumors in which low absolute increases in tumor size are interpreted as a substantial percentage difference (9). Conversely, the explosion of PD is not escalated depending on the growth rate and changes over time. Although the definition is necessarily arbitrary for comparison purposes, most glioblastomas present with mixed treatment response (admixture of TP and PsP or regional heterogeneity with TP in one area and PsP in another); for instance, classifying patients with glioblastoma as PsP considering the overall change in size when some areas are in fact progressing result in missing the early turning point for therapeutic intervention.

mRANO criteria use the postradiation time point as the baseline scan and require a minimum period of 12 weeks after the completion of CCRT and a repeat scan after four weeks for confirmation of TP/PsP unless the site of PD is distant from the radiation field or there is pathologic evidence of TP (5). Noteworthy, immediate postsurgical residual volume is an important prognostic factor, reducing OS in glioblastoma, and MRI, mainly within 72 hours postsurgery, is warranted to reduce confusion with postoperative reactive enhancement. mRANO criteria primarily rely on highly unpredictable, transient radiographic changes that often accompany the initial chemoradiation phase to use the postradiation MRI as the baseline scan. However, PsP can be present beyond 12 weeks (up to six months), and these first 12 weeks are a critical phase in which at least 50% of patients experiencing radiographic changes have true progression and are prevented from receiving alternative therapies for recurrent disease, thereby adversely affecting eventual outcomes (18).

The time for intervention can be even longer in patients with glioblastoma undergoing immunotherapy, assessed by the iRANO criteria. According to iRANO criteria, patients presenting with a lesion at the site of the original tumor, with a probable diagnosis of TP, or even showing new lesions at distant locations within the first six months of the onset of immunotherapy but clinically stable should continue with current immunotherapy. Furthermore, per iRANO, confirmation of radiographic progression can be performed at 12 weeks follow-up MRI instead of four weeks (19), a relative drawback, considering that patients with GBM have a low median OS and may be potentially receiving noneffective immunotherapy. In addition, trusting the patient's symptoms for changes in clinical management, while generally helpful, can also be misleading, as patients with a smaller TP or TP in nondominant locations will likely be relatively asymptomatic, preventing early institution of appropriate therapy, which may be surgical resection before too late (i.e., tumor spreading across corpus callosum or involving more eloquent regions and thereby becoming unresectable). Therefore, there is a pressing need to develop robust and quantitative imaging biomarkers which are “therapy agnostic” for reliably assessing “real-time” tumor dynamics and redefining the treatment response in patients with glioblastoma.

Noteworthy, utilizing physiologic and metabolic imaging parameters have been investigated to accurately determine treatment outcomes and define clinically beneficial endpoints, including neurologic and immunologic functions. In gliomas, cellular density and tumor grade are directly related to the degree of water restriction on diffusion-weighted imaging (DWI) with low apparent diffusion coefficient (ADC) values. In addition, ADC has shown value in patients treated with antiangiogenic agents and immunotherapy (20, 21). Specifically, minimum ADC values from enhancing areas could differentiate between inflammation and progressive tumor in patients treated with dendritic cell vaccine (22). The pathologic characteristics of tumor vasculature assessed by DSC–PWI can also help to discriminate between TP and PsP. In TP, there is a marked increase in angiogenesis that leads to an increase in relative cerebral blood volume (rCBV). In contrast, PsP is characterized by an increased inflammatory response, local accumulation of edema, and abnormal vascular permeability leading to decreased rCBV (23, 24).

In GBMs, spatial and temporal intratumoral heterogeneity can result in a mismatch in the findings from diverse neuroimaging parameters. Therefore, the use of a single imaging technique or parameter may not always be reliable for evaluating treatment response. Integrating the unique strengths of different imaging techniques, such as DTI and DSC–PWI, in a multiparametric approach showed a precise assessment of treatment response (25). A meta-analysis (26) showed the potential surrogate endpoint of multiparametric MRI for the assessment of early-treatment response in newly diagnosed glioblastoma treated with CCRT, with pooled sensitivity and specificity of 84% and 95%, respectively. In addition, the multiparametric analysis could accurately evaluate treatment response to immunotherapies, such as immune checkpoint inhibitors (27), EGFR variant III (EGFRvIII) targeted CAR T (28), and IL4 receptor–targeted immunotherapy (29) in patients with recurrent glioblastoma.

MRS can also add utility to the response assessment of nonenhancing tumors, differentiation of TP versus PsP, and genetic profiling of gliomas. Elevated lipid and low choline (Cho)/N- acetyl aspartate (NAA) ratios have been reported in association with PsP (2). Whole-brain echo-planar spectroscopic imaging (EPSI) following standard therapy also showed higher Cho/Creatine (Cr) and Cho/NAA ratios in patients with TP compared with PsP. EPSI can detect “invisible” tumor cells occult in conventional MRI sequences, which is useful for planning the course of treatment for these patients, such as optimizing the extent of resection and the dose and target area of radiation (2, 30). Several amino acid(aa)-based PET imaging tracers have emerged as alternative candidates for metabolic imaging of brain tumors. These tracers are characterized by high tumor-to-brain contrast based on their relatively high specificity for neoplastic cells and low accumulation in normal brain tissues (31).

Despite promising findings, metabolic and physiologic imaging techniques are associated with certain shortcomings. It is well known that even minor differences in hardware components or sequence parameters may result in significant changes in image contrast and signal intensity values, hindering the interpretation of imaging results from different treatment centers. Therefore, widespread adoption of these imaging techniques into routine clinical workflow requires standardization, harmonization of data acquisition, and processing protocols, along with the application of well-defined quality assessment/control procedures. Fortunately, consensus guidelines have been proposed to implement diffusion, perfusion, spectroscopy, and PET imaging techniques across different clinical sites (32–35). Additional improvements in this field require data sharing and large multicentric validation studies.

AI is an emerging translational field that holds promise to improve the precision of diagnostic and therapeutic methods. Radiomic, radiogenomic, and radiopathomic tools provide a means of noninvasive sampling of tumor microenvironments, allowing for a dynamic and comprehensive evaluation of regionally heterogeneous brain tumors (36–38). Although using machine learning–based models in distinguishing TP from PsP, some studies have demonstrated that multiparametric (anatomic MRI, ADC, and CBV) radiomics model perform significantly better than a monoparametric radiomics model using either conventional MRI, DWI, or PWI-derived parameters (39). These findings further emphasize the importance of applying multi-parametric approach even if we use more sophisticated machine learning–based models in comparison to traditionally used logistic regression models.

However, AI methods have been challenged by insufficient training, heterogeneity of imaging protocols across hospitals, and lack of generalization to new patient data. These challenges prompted the development of the ReSPOND (Radiomics Signatures for PrecisiON Diagnostics) consortium for glioblastoma [10 institutions, 3 continents, more than 3,300 de novo patients with GBM, and datasets from The Cancer Imaging Archive (TCIA)] to further develop and test AI-based biomarkers by moving from single-institution studies to generalized, well-validated predictive biomarkers (40). Notable, AI-based concepts are in continuous development but not fully integrated into daily practice. We, at this time, still depend on the combination of humans and computers (augmented intelligence) to incorporate information from images, AI tools, and electronic medical records to enhance the stratification of patients into more precise therapeutic pathways enabling dynamic treatment monitoring in this era of personalized medicine (41).

Beyond histologic and radiomics diagnosis, there is increasing recognition of integrating an objective assessment of molecular information. The integration of molecular data, such as MGMT methylation status, EGFRvIII, and isocitrate dehydrogenase (IDH) mutation, can further improve response assessment. It is well recognized that the methylation status of MGMT and IDH mutated gliomas are independent favorable prognostic factors in patients with GBM, and MGMT methylated patients are more frequently associated with PsP (42, 43). In addition, EGFRvIII has been explored extensively for targeted therapies, more recently in CAR T-cell immunotherapy (44). In the era of radiogenomics, imaging signatures of tumor genomics can help in identifying the patients who benefit from certain targeted therapies. Specifying molecular properties of gliomas and predicting their changes over time can also optimize treatment surveillance.

In conclusion, there is an urgent need for a reproducible, quantitative, and objective model to accurately evaluate treatment-related temporal changes in “real-time”, particularly in the early posttreatment window, to improve consistency and reduce variability in assessing outcomes in neuro-oncology. The quantitative, multiparametric neuroimaging approach substantiating the well-established mRANO criteria holds promise to further improve accuracy and automation in the future (Fig. 1).

No disclosures were reported.