Yae Won Park¹, Ki Sung Park², Sung Soo Ahn¹, Inho Park¹, and Se Hoon Kim^1
1Yonsei University College of Medicine, Seoul, Korea, Republic of, ²Pohang University of Science and Technology, Seoul, Korea, Republic of

CDKN2A/B homozygous deletion is a key molecular marker in the 2021 WHO classification of IDH-mutant astrocytomas; presence of CDKN2A/B homozygous deletion results in a grade of 4.   On multivariable analysis, infiltrative pattern, larger tumor volume, and higher 95th percentile of nCBV were independent predictors of CDKN2A/B homozygous deletion.  The infiltrative pattern and larger volume may be reflected from the increased tumor cell proliferation, whereas the higher 95th percentile of nCBV may reflect the increased angiogenesis due to CDKN2A/B homozygous deletion.

Jacob Ellison^1,2,3, Nate Tran^1,2,3, Julia Cluceru^1,2, Joanna Phillips^4,5, Anny Shai⁵, Devika Nair¹, Annette Molinaro⁵, Valentina Pedoia^1,2,3, Yan Li^1,2,3, Javier Villanueva-Meyer¹, Mitchel Berger⁵, Shawn Hervey-Jumper⁵, Manish Aghi⁵, Susan Chang⁵, and Janine Lupo^1,2,3
1Radiology and Biomedical Imaging, UCSF, San Francisco, CA, United States, ²Center for Intelligent Imaging, UCSF, San Francisco, CA, United States, ³Graduate Group in Bioengineering, UCSF - UC Berkeley, San Francisco, CA, United States, ⁴Brain Tumor Research Center, UCSF, San Fransisco, CA, United States, ⁵Neurological Surgery, UCSF, San Francisco, CA, United States

Treatment-induced effects can mimic tumor recurrence and pose a challenge to accurately assessing treatment response. We aim to provide a machine learning framework and identify important imaging features for discriminating treatment-induced injury from recurrent glioblastoma at biopsy level resolution. Our best model performs with a mean AUC of .77+/-0.11 across 4 fold cross-validation of 108 tissue samples. rCBV, choline-to-NAA index (CNI), and normalized lipid levels were the top three most import features in distinguishing treatment effects from recurrent tumor.

Sara Merkaj¹, Khaled Bousabarah², Lin MingDe¹, Andrej Pala³, Gabriel Cassinelli Petersen¹, Leon Jekel¹, Ryan Bahar¹, Niklas Tillmanns⁴, Ajay Malhotra¹, Malte Westerhoff², and Mariam Aboian^1
1Yale School of Medicine, New Haven, CT, United States, ²Visage Imaging, Berlin, Germany, ³Ulm University, Ulm, Germany, ⁴University of Düsseldorf, Düsseldorf, Germany

Tumor segmentation is a laborious process, which impedes the progress of data production for development of classification/prediction algorithms. We present a novel PACS-based workflow for deep learning-based auto-segmentation of gliomas that allows generation of annotated images during clinical workflow. We developed a U-Net auto-segmentation algorithm natively imbedded in PACS and trained on BraTS dataset. Subsequent retraining on tertiary hospital dataset was performed and generation of new segmentations, allowing labeling of 440 gliomas in a three-months period. This novel approach for annotated data generation allows real-time building of large, labeled datasets by experts in the field.

Adrian Lazaro Breto¹, Kaylie Cullison², Kolton Jones³, Olmo Zavala-Romero⁴, John C Ford¹, Eric A Mellon¹, and Radka Stoyanova^1
1Radiation Oncology, University of Miami Sylvester Comprehensive Cancer Center, Miami, FL, United States, ²University of Miami Miller School of Medicine, Miami, FL, United States, ³West Physics, Atlanta, GA, United States, ⁴Department of Earth, Ocean, and Atmospheric Sciences, Florida State University, Tallahassee, FL, United States

Identifying early progressors following treatment for Glioblastoma (GBM) is paramount in GBM management. MRI-RT platforms provide opportunity for daily MRI of patients. We identified early changes in tumor volume typically starting week 3 or 4 of treatment. We hypothesize tumor volume kinetics are associated with outcome and allow for adapting treatment. We develop a deep learning solution for automatic volume delineation on daily scans, allowing real time monitoring of tumor changes and reducing time burden of segmentation. We obtained DSC for tumor lesion and resection cavity on training and test datasets (mean±standard deviation) 0.87±0.128 and 0.9±0.122; 0.74±0.233 and 0.8±0.277, respectively.

Aaron Rulseh¹ and Josef Vymazal^1
1Dept. of Radiology, Na Homolce Hospital, Prague, Czech Republic

Glioblastoma (GBM) is the most common and aggressive primary brain tumor. Any methods that may improve confirmation or early detection of progression are highly desirable. T2 relaxometry shows great promise in the monitoring of GBM patients following complex therapy. We found WB metrics, such as median 1/T2, correlated with PFS and were able to distinguish progression from pseudo-progression. A deep neural network trained on automatically segmented 1/T2 data achieved an F1 accuracy of 85% in classifying progression with a threshold of 18 months. 

Laiz Laura de Godoy¹, Sumei Wang¹, Shadi Asadollahi¹, MacLean P. Nasrallah¹, Donald M. O’Rourke¹, Steven Brem¹, Arati Desai¹, Laurie Loevner¹, Suyash Mohan¹, and Sanjeev Chawla^1
1Perelman School of Medicine at the University of Pennsylvania, Philadelphia, PA, United States

The highly aggressive nature of glioblastoma (GBM) leads to a dismal prognosis, and alternative therapies are being sought. Immunotherapy, such as the Dendritic Cell (DC) vaccine, is currently being evaluated for recurrent GBMs in clinical trials. Follow-up imaging after this immunotherapeutic approach often mimics disease progression on conventional images, making standard MRI evaluation challenging. Our findings indicate that our previously established multiparametric MRI-based prediction model has the capacity to accurately assess response to DC vaccine in recurrent GBM patients, objectively characterizing as either TP or PsP GBM patients, avoiding interruptions in satisfactory treatments, and preventing invasive procedures in PsP cases.

Zahra Riahi Samani¹, Drew Parker¹, Steven Brem², and Ragini Verma^1
1DiCIPHR (Diffusion and Connectomics in Precision Healthcare Research) Lab, University of Pennsylvania, Philadelphia, PA, United States, ²Department of Radiology, Department of Neurosurgery, University of Pennsylvania, Philadelphia, PA, United States

We developed a novel, microstructure-based, voxel-wise map of the peritumoral region in glioma brain tumors using DTI-based free water volume fraction map and deep-learning. This novel map captures the infiltrative heterogeneity of peritumoral region and can differentiate between gliomas with distinct IDH1 mutation status (IDH-mutant vs. IDH-wildtype). Thus, this new derived map that incorporates microstructure information can be used as a new diffusion based radiomic feature for various oncological investigations involving mutation status.

Virendra Kumar Yadav¹, Rakesh Kumar Gupta², Sumeet Agarwal³, and Anup Singh^1,4
1Center for Biomedical Engineering, Indian Institute of Technology, Delhi, India, ²Fortis Memorial Research Institute, Gurugram, India, ³Electrical Engineering, Indian Institute of Technology, Delhi, India, ⁴Biomedical Engineering, All India Institute of Medical Sciences, Delhi, India

Precise safe surgical resection and precisely directed radiation are important in clinical practice for low-grade gliomas (LGGs) patients in order to minimize the neurological deficit and radiation toxicity, respectively. Clinicians find difficulty in defining the border between the non-enhancing tumor and edema components. In this study, machine learning-based models were developed in order to distinguish non-enhancing tumors from vasogenic edema using quantitative perfusion parameters obtained using dynamic-contrast-enhanced MRI. The proposed approach may help in assisting radiologists, by defining precise tumor boundaries and hence, results in improving patients’ quality of life and overall survival.