Hi, everyone. Thanks for taking the time to view my presentation, and thank you to the session organizers for coordinating virtual presentations during this unprecedented time. My name is Elie Veliani. I'm a second-year medical student at the Icahn School of Medicine at Mount Sinai, and I'll be discussing some of my recent work with Dr. Hadjipanais and Dr. Owerman, titled Glioblastoma MRI Analytics Using Deep Learning. As a brief introduction, glioblastoma is the most common malignant central nervous system tumor. It accounts for roughly 15% of brain tumors, has a median survival rate of less than 15 months, and a five-year survival rate of less than 10% with aggressive treatment. Given this, intense research is underway to identify biomarkers of response with the hope of informing patient prognoses and improving clinical management in the form of personalized therapies. Genetic markers that provide insight into tumor progression and outcome have previously been identified. For example, IDH mutant glioblastomas exhibit improved survivorship relative to IDH wild-type glioblastomas, and glioblastomas with MGMT promoter methylation demonstrate increased susceptibility to radiotherapy in combination with temozolomide as opposed to radiotherapy as a single agent. However, the challenge is that tumor characterization in the clinical setting relies on invasive biopsies or craniotomies that come with their own risks and are limited to easily accessible tumors. High-grade gliomas, for example, are known to infiltrate into peritumoral regions that often may be surgically inaccessible. As such, the goal of our work was to develop an end-to-end machine learning pipeline that ingests a patient's preoperative MRI, segments tumor regions of interest, and subsequently predicts genetic characteristics. In order to conduct this work, we sought to obtain data that contained preoperative MRIs of gliomas and associated genomic information characterizing tumor status. We utilized the publicly available Brain Tumor Segmentation dataset to conduct our work. It consists of 411 preoperative full-series MRIs of low-grade and high-grade gliomas derived from 19 institutions. The data also comes with patient demographic information and, most important for algorithm development, gold-standard tumor segmentations verified by neuroradiologists. Next, we mined the Cancer Genome Atlas to match each patient MRI in the aforementioned dataset with relevant genomic information to characterize tumor status. And the resulting matched dataset contained 119 low-grade gliomas and 292 high-grade gliomas, with approximately 41% of those cases having IDH1 mutations, 32% with MGMT promoter methylation, 11% with 1p19q code deletions, and 11% with EGFR amplification. Our algorithm pipeline consisted of two steps. In the first step, we developed a segmentation algorithm to ingest preoperative brain MRIs and output tumor regions of interest. And in the second step, we cropped the tumor region of interest and fed it to a neural network classifier to identify tumor status. When evaluating the efficacy of segmentations, it's important to do so both qualitatively and quantitatively. Here's a randomly chosen T1 MRI that our algorithm was asked to segment. You can see the tumor show up right there. In the middle is our model segmentation, and on the right is the neuroradiologist segmentation for comparison. As you can see, visually, the algorithm does a nice job of segmenting out the tumor core and surrounding edematous regions. We quantitatively evaluated the quality of our model segmentations using the DICE coefficient, which determines the percent of voxels that were correctly classified. Based on this metric, our algorithm identified 89% of the whole tumor and tumor core and approximately 84% of the enhancing tumor, which indicates that our model is segmenting the tumor with granularity that's on par with current state-of-the-art approaches. Having verified the efficacy of our segmentations, we'll now evaluate the second aspect of our pipeline, which involves biomarker prediction. Area under the curve denoted AUC is the standard metric for evaluating predictive performance. Results are in this table. I know there's a lot here, so I'll point out the important findings. First, our baseline algorithm was able to differentiate tumor grade at an AUC of 97%. That same algorithm, when pre-trained on tumor grade and subsequently tuned to tumor status, yielded the best performance in predicting IDH status with an AUC of 93%. Incorporating patient demographic information such as age and sex into the model boosted prediction of tumor grade and yielded the best performance in predicting 1p19q co-deletion and EGFR amplification with AUCs of 83% and 71% respectively. This is interesting because it indicates that patient demographics may provide additional signal in biomarker characterization. Finally, the combination of pre-training and demographics yielded the best performance among all our experiments in predicting MGMT methylation status at an AUC of 77%. In conclusion, our work shows that it's possible to characterize glioma biomarkers from preoperative MRIs with high efficacy. We also present a novel multimodal training approach that incorporates MRIs in patient demographics to enhance efficacy. But our work is not complete. Future work will entail incorporating additional radiomics characteristics such as tumor location and size to determine whether additional improvements to performance are possible. We'll also need to externally validate our pipeline and prospectively assess its efficacy in clinical workflows to determine its utility in a real-world clinical setting. Lastly, because these algorithms are currently a black box, we'd like to better understand the radiomics features that add the greatest signal in biomarker characterization to improve our understanding not only of how these algorithms make decisions, but also our understanding of clinical features that impact patient prognoses. I'd like to thank my mentors, Drs. Hajipanais and Overman for their guidance, my peers at Mount Sinai and the AI Sinai Consortium for their support, and for the Neurosurgery Research and Education Foundation for funding this work. Thank you.

glioblastoma

MRI analytics

deep learning

genetic characteristics

tumor segmentation