Gender-Specific AI Model Advances Understanding of Glioblastoma Progression and Treatment Response

Despite advances in the management of primary glioblastoma ― chemotherapy, radiation therapy and surgery ― the disease continues to have a poor prognosis with patients typically surviving just 15-18 months. Clinicians face numerous challenges when caring for the 15,000 Americans diagnosed with glioblastoma each year.

Among the most difficult clinical problems is distinguishing on current imaging between disease progression and pseudo progression, particularly following temozolomide (Temodar) treatment, due to inflammatory changes that occur in the brain after treatment with chemotherapy and radiation. Benign treatment-related radiation effect and “true” tumor recurrence mimic each other clinically and radiographically. Pseudo progression occurs in some 40 percent of patients, with most of them harboring the O6 -methylguanine-DNA methyl-transferase methylated tumors. The result is that a large number of patients with benign chemoradiation-related changes often undergo additional imaging, and sometimes unnecessary and invasive intra-cranial surgery or biopsies.

Evolving research is proving valuable in distinguishing radiation effects from tumor recurrence. Building upon an earlier study from our group that suggested gender differences should be considered an influence in prognostic outcome¹ and a more recent study to identify signaling pathways that drive sex-specific tumor biology and treatment²,, we are moving ahead in taking the next steps to generate what we believe will be powerful and more accurate tools in risk-stratifying glioblastoma patients for personalized decision-making.

By mining and extracting computational features that are not visually apparent to radiologists or clinicians, it is possible to map and distinguish between radiation-related changes and tumor progression. Funded by an R01 NIH grant (1R01CA264017-O1A1), our group is working on developing a “sex-specific” Image-based Recurrence Risk Classifier (IRRisC), that will employ advanced radiomics and machine learning approaches to distinguish radiation effects and tumor recurrence, for men and women. Uniquely, unlike the “black box” machine learning and deep learning approaches that have previously been explored in the literature, IRRisC will leverage “hand-crafted” image features that capture the underlying disease heterogeneity extant in GBM tumors, via measurements of local gradient entropy (degree of disorder being associated with aggressiveness), as well as a new class of biophysical deformation attributes and surface topology ― features that capture the tumor microenvironment on routine MRI (Gd-T1w, T2w, FLAIR) sequences.

Our results so far on preliminary analysis have been very promising with an accuracy of close to 90% in distinguishing histologically proven radiation necrosis from tumor recurrence, on a cohort of multi-institutional datasets. Our goal is to demonstrate value of IRRisC as a decision support (to complement neuroradiologists in decision making) on a much larger cohort of histologically confirmed cases of radiation necrosis and tumor recurrence via our multi-institutional collaborations involving Case Western, Miami Cancer Institute, University of Wwashington-Madison, Northwestern University and Cleveland Clinic. The collaboration will also include GE Research, through which we will be extending these diagnostic tools to be deployed over a cloud-platform. The cloud-platform will provide IRRisC a global reach and accessibility, both within the clinics and hospitals within the U.S. as well as internationally. In addition, the research has implications for immunotherapeutic therapies for glioblastoma, which also may cause inflammation in the brain that is not actually indicative of aggressive tumor growth.

As we lead the charge in developing these improved diagnostic tools in collaboration with GE Research, we are hopeful that this five-year initiative will enable us to move into clinical trials and, ultimately, toward improving quality of life of GBM patients.

_REFERENCES:

_{1. Gender-specific Probabilistic Atlases of Glioblastoma Reveal Impact of Tumor Location on Progression Free Survival. https://doi.org/10.1093/neuonc/noz175.333. Published Nov. 11, 2019.}

_{2. Sexually dimorphic radiogenomic models identify distinct imaging and biological pathways that are prognostic of overall survival in glioblastoma. https://doi.org/10.1093/neuonc/noaa231. Published Feb. 25, 2021.}