Brain-Dynamics: A Unified Platform for Longitudinal Tracking and Quantitative Assessment of Brain Lesion Dynamics In Neuroimaging
Abstract
Purpose
To address fragmentation and variability in longitudinal brain lesion assessment, we developed Brain-Dynamics, a vendor-neutral platform integrating auto‑segmentation, multimodal co‑registration, lesion labeling/tracking, and quantitative analytics for research and clinical use.
Methods
Brain-Dynamics connects to PACS/EHRs, supports MRI/CT image analysis across timepoints, enables interactive contour editing, and exports DICOM for clinical integration. Three AI-based auto-segmentation models were built on nnU-Net and trained/validated on retrospective stereotactic images acquired from vestibular schwannomas (VS), brain metastases (BMs), and glioma patients, respectively. Using the Montreal Neurological Institute (MNI) space as a standard atlas, multimodal images are rigidly co-registered via mutual information (MI) and lesions names are standardized with their laterality and lobes. The platform supports quantitative measurements, including volume, three principal diameters, BMs velocity and multimodality radiomic analysis.
Results
Brain-Dynamics’ performance was evaluated on the accuracy of lesion detection, segmentation, and labelling. VS segmentation evaluated on 20 patients achieved Dice of 0.90±0.05, approaching inter-observer variability. Longitudinal VS volume tracking showed a consistent pattern of post‑SRS effects: transient enlargement due to treatment‑related edema followed by stabilization or shrinkage indicating tumor control. BMs detection and segmentation evaluated on 79 patients with a total of 536 lesions achieved sensitivity of 89.70±1.85%, precision of 97.34±0.77%, Dice of 0.92±0.06, and lesion labelling accuracy of 100%. BMs velocity and volume tracking were used to facilitate BMs re-SRS studies. Glioma segmentation on 840 image sessions (T1c+FLAIR) achieved Dice of 0.77±0.29, 0.84±0.23, and 0.86±0.12 for tumor core, enhancing tumor, and edema, respectively, despite large heterogeneity in tumor shapes and intensities. The segmented substructures were used to facilitate studies of their morphological relationship and progression.
Conclusion
Brain-Dynamics supports AI auto-segmentation with longitudinal lesion tracking and analysis on three brain diseases. Future work will expand the platform’s application to other brain indications and prospectively evaluate its impact on clinic workflow and treatment decision-making.