Artificial Intelligence-based decision support for chemotherapy-response assessment in Brain Tumors

Wm S. Middleton Memorial Veterans Hosp, Madison WI

Investigators

Linked publications & trials

Abstract

ABSTRACT: In 2020, over 23,000 patients in the US will be diagnosed with Glioblastoma (GBM), a highly aggressive brain tumor, with a dismal median survival of 15-18 months. Studies focusing on Gulf War Veterans especially those exposed to nerve agents in Iraq in 1991 have shown a higher risk of brain tumors among neurological diseases and a distinct neurological brain pattern as compared with the other Veterans. The standard-of-care for GBM consists of surgical resection followed by radiotherapy combined with concomitant and adjuvant chemotherapy. However, ~50% of GBM patients do not respond favorably to chemoradiation following surgery. A priori identification of non-responders could allow for selection of these patients as potential candidates for genomically-driven drug therapies (over 64 ongoing clinical trials in the US) over conventional treatment. Further, chemotherapy costs >$100K/year. There is hence an unmet need to develop and validate predictive biomarkers to identify up front which Veteran patients will not benefit from chemotherapy. Another significant challenge in GBM management is the differentiation of suspicious lesions on post-treatment MRI, as tumor recurrence or treatment-induced radiation effects. In the absence of reliable diagnosis, patients with a benign treatment effect have to undergo an unnecessary surgical confirmation biopsy. The co-morbidities due to unnecessary biopsies disproportionately impact Veteran GBM patients who tend to be older and have increased comorbidity burden. Consequently, developing a companion diagnostic solution using clinical MRI could represent a compelling solution in substantially improve quality-of-life years for Veteran GBM patients by sparing them of the side-effects of surgery, while providing timely management in patients with tumor recurrence. Recently, we have developed a new “Neuro-Image Risk Classifier” (NeuRisC), that uses artificial-intelligence (AI)-driven computational features corresponding to the micro-architectural measurements of disorder in the local intensity gradients (i.e. gradient entropy) on Gadolinium (Gd)-T1w MRI; the initial version of NeuRisC has been shown to (a) be prognostic of GBM survival on n=203 studies (p<0.001), and (2) have an accuracy of 85% (a 37% improvement over expert readers) on n=58 studies in distinguishing radiation effects from tumor recurrence. In this VA project, we propose to further improve, and validate the accuracy of NeuRisC by expanding our initial feature set (using Gd-T1w MRI alone) by including (1) additional features from anatomical (T2w, FLAIR) and functional MR sequences (perfusion), (2) a new class of biophysical deformation attributes from “normal” brain parenchyma, and (3) peritumoral features from outside the lesion. In Aim 1, we will develop (NeuRisC)predict as a predictive image-based marker of benefit to chemotherapy by combining intra- and peri-tumor gradient entropy and biophysical deformation attributes from “normal” brain parenchyma. Similarly, (NeuRisC)diagnose will be developed in Aim 2 by including lesion and peri-lesional features from pre- and multiple post-treatment MRIs, to improve discrimination of radiation effects and tumor recurrence. Overcoming limitations in previous work pertaining to small cohorts & lack of spatially mapped ex-vivo histology, NeuRisC modules will be validated on a large multi-institutional cohort of >1000 studies with co-localized tissue and MRI scans obtained across multiple biopsies/lesion. This cohort will also allow for establishing associations of NeuRisC features with underlying histological/molecular tumor characteristics - a prerequisite for clinical adoption. Lastly, NeuRisC modules will be deployed at Northeast Ohio & Tennessee VA Healthcare Systems to validate their utility as decision support. On an independent cohort of N=200 MRIs from Veteran patients, interpretation results from oncologists and radiologists at these two VA sites will be compared, with and without NeuRisC, to evaluate added benefit of NeuRisC as decision support. Criteria for success will be to demonstrate that NeuRisC is able to (a) predict GBM patients that respond favorably to chemoradiation with >90% accuracy, and (b) is non-inferior to accuracy for invasive biopsies (85-90% accuracy), thereby avoiding biopsies in patients with a benign radiation effect.