3628 - Machine Learning with Connectomics to Predict Prognosis in Glioblastoma after Radiation
Presenter(s)
V. Gopalakrishnan1, V. Parekh2, M. C. LeCompte1, E. Huang1, A. Suresh3, Y. Tarui3, A. Li4, J. M. Reyes5, K. J. Redmond1, M. A. Jacobs2, and L. R. Kleinberg1; 1Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 2Russell H. Morgan Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD, 3Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD, 4Department of Psychology, Johns Hopkins University, Baltimore, MD, 5Department of Neuroscience, Johns Hopkins University, Baltimore, MD
Purpose/Objective(s): Tumor connectomics is a novel MRI-based complex graph theory framework that describes network relationships within tumor and surrounding tissues. In this study, we designed a 4D-tumor-connectomics framework for topological characterization of brain tumors in multiparametric Magnetic Resonance Images (mpMRI). We then integrated 4D-tumor-connectomics with clinical data to differentiate survival outcomes in glioblastoma (GBM), a highly aggressive brain cancer with variable prognosis.
Materials/Methods: Patients with IDH-wildtype GBM who received maximally safe resection followed by conventionally fractionated radiation therapy (RT) with concurrent temozolomide were included. We curated a patient cohort of 25 patients (11 patients with overall survival (OS) < 1 year from end of RT versus 14 patients with OS > 2 years). Multiparametric MRI sequences (T1-weighted pre- and post-contrast, FLAIR, DWI, ADC) were obtained preoperatively, postoperatively (< 72 hrs post-surgery), at first post-RT follow-up, and at first progression per RANO 2.0. Not all patients reached progression before death. Tumor volumes of interest were manually segmented at each time point by trainees and verified by radiation oncologists. For each patient, 88 connectomics features were extracted from the mpMRI sequences. An additional 20 clinical/treatment variables, including age, performance score, MGMT methylation status, and presence of radiation necrosis, were collected. An Integrated Radiomics Informatics System (IRIS) based on an Isomap support vector machine (IsoSVM) model with nnU-net was used to distinguish OS differences using leave-one-out cross validation. Analysis was performed on both the clinical data and time-series image data for each patient.
Results: The IsoSVM model achieved a sensitivity of 0.91, specificity of 0.86, and an AUC of 0.98 in predicting OS < 1 year based on tumor topological characteristics. The model maintained an AUC of 0.98 across all time points except post-RT progression, where the AUC was 0.93. Tumors in patients with OS < 1 year were highly connected, exhibiting significantly higher median degree centrality (p = 0.04) and clustering coefficient (p = 0.05), while having significantly lower betweenness centrality (p = 0.02).
Conclusion: Our results highlight the importance of understanding and characterizing tumor topology to assess treatment effects and predict patient survival. Tumors in patients with OS < 1 year exhibited denser, more clustered, and less centralized networks. This could imply a diffusely infiltrative growth pattern, where no single node (region) controls the tumor, making surgical resection and focal therapies less effective. This architecture may facilitate rapid progression, treatment resistance, and efficient intratumoral signaling, contributing to poor prognosis. Post-RT changes in tumor connectomics metrics suggest that RT alters tumor architecture.