Data-Driven Simulation of Fisher-Kolmogorov Tumor Growth Models Using Dynamic Mode Decomposition.

Area of Science:

• Computational oncology
• Biophysics
• Machine learning

Background:

• Personalized cancer treatment relies on patient-specific computational models for predicting tumor growth and treatment response.
• High computational costs of complex spatiotemporal cancer models hinder clinical applicability.
• Dynamic-mode decomposition (DMD) is an unsupervised machine learning technique for system analysis and prediction.

Purpose of the Study:

• To utilize DMD for constructing low-dimensional representations of cancer models to accelerate simulations.
• To apply DMD to Fisher-Kolmogorov models for personalized tumor growth and treatment response forecasting.
• To reduce the computational burden of patient-specific cancer simulations.

Main Methods:

• Application of dynamic-mode decomposition (DMD) to Fisher-Kolmogorov models.
• Development of low-dimensional representations for organ-scale biomechanistic cancer models.
• Validation of DMD accuracy and error bounds over clinically relevant parameter spaces.

Main Results:

• DMD implementation achieved short- to medium-term prediction errors under 1% and long-term errors under 20%.
• Accurate and bounded-error reconstructions were obtained even with short training periods.
• DMD successfully reconstructed tumor-induced host tissue deformation fields.

Conclusions:

• DMD offers a data-driven approach to significantly reduce computational overhead in personalized cancer simulations.
• This method facilitates faster tumor forecasting, parameter identification, uncertainty quantification, and treatment optimization.
• The proposed approach enhances the clinical actionability of computational oncology tools.