Physics-Regularized Multi-Modal Image Assimilation for Brain Tumor Localization

Authors: Michal Balcerak, Tamaz Amiranashvili, Andreas Wagner, Jonas Weidner, Petr Karnakov, Johannes C. Paetzold, Ivan Ezhov, Petros Koumoutsakos, Benedikt Wiestler, bjoern menze

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate enhanced coverage of tumor recurrence areas using real-world data from a patient cohort, highlighting the potential of our method to improve model-driven treatment planning for glioblastoma in clinical practice.
Researcher Affiliation Academia 1University of Zurich, 2Technical University of Munich 3Harvard University, 4Imperial College London
Pseudocode No The paper describes procedures and methods in paragraph text but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes *Code is available at https://github.com/m1balcerak/PhysRegTumor
Open Datasets Yes We use a publicly available dataset of 58 patients with preoperative MRI scans and preoperative FETPET imaging, along with follow-up MRI scans at the time of the first visible tumor recurrence [22].
Dataset Splits No The paper mentions using synthetic data for hyperparameter tuning ('All α parameters were determined based on experiments with synthetic data') and generating '100 synthetic patients' but does not specify exact train/validation/test split percentages or sample counts for this dataset for validation.
Hardware Specification Yes Using the Adam optimizer, convergence takes around 3 hours on an NVIDIA RTX A6000 GPU.
Software Dependencies No The paper mentions using the 'Adam optimizer' but does not specify any software names with version numbers for libraries, frameworks, or programming languages used in the implementation.
Experiment Setup Yes Our computational model uses a multi-resolution method [37] with additional, coarser grids to accelerate convergence. We employ four levels of grid refinement, resulting in 152,880,048 unknowns. At the finest grid level of 72 x 72 x 72 x 96, each grid point contains four unknowns: three for particle positions and one for tumor cell density. ... Using the Adam optimizer... ...β modulates the steepness of the transition and is set to β = 50. ...α offsets the thresholds and is set to α = 0.05.