2533 - Brain-Dynamics: Toward Standardized Longitudinal Brain Lesion Analytics for Precise and Adaptive Neuro-Oncology
Presenter(s)
M. Chen1, G. A. Szalkowski2, R. Prashad2, J. S. Fernandes3, F. Lam4, Q. Wang1, Y. Zhu1, S. G. Soltys2, E. L. Pollom2, E. Rahimy2, I. C. Gibbs2, D. Park4, Y. Hori4, R. D. Timmerman5, T. Dan5, Z. Wardak5, M. Dohopolski5, J. B. De Vis5, H. Jiang1, X. Gu2, and W. Lu5; 1Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Dallas, TX, 2Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA, 3Department of Radiation Oncology, Stanford University, Stanford, CA, 4Department of Neurosurgery, Stanford University School of Medicine, Stanford, CA, 5Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX
Purpose/Objective(s):
Longitudinal assessment of brain lesions remains fragmented, labor-intensive, and inconsistent across clinical and research workflows. We developed and validated Brain-Dynamics, a vendor-neutral, web-accessible platform that integrates with clinical imaging systems to identify, segment, label, and longitudinally track brain lesions, enabling quantitative, reproducible, scalable evaluation of brain lesion dynamics.Materials/Methods:
Brain-Dynamics integrates with PACS/EHR systems to provide end-to-end longitudinal lesion assessment, including automated lesion identification and segmentation, clinician-guided contour editing, standardized lesion naming, and export of results as DICOM objects, and supports MRI and CT analysis. A front-end web client (Rust/WebAssembly) enables visualization, editing, and task management, while a back-end server executes segmentation, co-registration, and tracking tasks. Three disease-specific nnU-Net–based models were developed and validated for vestibular schwannoma (VS) segmentation, brain metastases (BMs) detection/segmentation with automated lesion labeling, and glioma tumor and substructure segmentation. Multimodal and longitudinal image alignment are performed via rigid co-registration to MNI space using mutual information, and lesions are automatically labeled by laterality and lobe, supporting consistent lesion identity across timepoints and longitudinal analysis. Quantitative outputs include lesion volume, principal diameters, growth trajectory, and multimodal radiomic features. DICOM export with standardized metadata ensures clinical interoperability and seamless workflow integration.Results:
Performance was evaluated across three indications. VS segmentation (n=20) achieved Dice 0.90±0.05, approaching inter-observer variability. Longitudinal VS tracking demonstrated characteristic post-SRS dynamics with transient enlargement followed by stabilization or shrinkage. BMs detection/segmentation (79 patients; 536 lesions) achieved sensitivity 89.70±1.85%, precision 97.34±0.77%, Dice 0.92±0.06, and 100% lesion labeling accuracy. Velocity and volumetric tracking facilitated re-SRS analyses. Glioma segmentation (210 patients; 840 sessions, T1c+FLAIR) achieved Dice 0.77±0.29 (tumor core), 0.84±0.23 (enhancing tumor), and 0.86±0.12 (edema), supporting substructure-based progression and survival modeling despite substantial heterogeneity.Conclusion:
Brain-Dynamics provides a unified, AI-enabled platform for automated segmentation, standardized lesion labeling, and longitudinal quantitative tracking across multiple brain diseases. By integrating automation with clinically meaningful analytics, the system accelerates workflow, improves reproducibility, and supports treatment-response research and data-driven decision-making. Ongoing work will expand disease coverage and prospectively evaluate clinical impact.