Physics-informed DynUNet for brain metastasis segmentation.

Area of Science:

• Neuro-oncology
• Artificial Intelligence
• Medical Imaging

Background:

• Detecting and segmenting small brain metastases is challenging for standard deep learning.
• Existing models lack biological information on metastasis growth, limiting performance on low-volume lesions.

Purpose of the Study:

• To investigate if integrating physics-informed (PI) tumor growth models into segmentation architectures can overcome size-dependent limitations in brain metastasis detection.
• To evaluate the effectiveness of PI-DynUNet in segmenting brain metastases of varying sizes and characteristics.

Main Methods:

• Developed PI-DynUNet by integrating a physics-based growth model with DynUNet.
• Trained and compared PI-DynUNet against U-Net variants on the BraTS-METS 2023 dataset using controlled settings.
• Evaluated performance across six lesion-size categories and assessed clinical context-specific performance using scenario-weighted Dice coefficients.

Main Results:

• PI-DynUNet demonstrated improved segmentation across all BraTS regions, with notable gains in non-enhancing tumor core (NETC) Dice (+5.3%).
• Optimal regularization weights (λ) were found to be tissue and lesion-size dependent.
• Context-dependent optimal models were identified, with PI-DynUNet (λ=0.01) excelling in enhancing-weighted scenarios and λ=1.0 showing superior balanced accuracy.

Conclusions:

• Physics-informed deep learning offers modest but significant improvements in brain metastasis segmentation.
• External validation confirmed the outperformance of PI-DynUNet over baseline DynUNet, with substantial gains in tumor-core Dice.
• Optimal PI-DynUNet configuration is application-specific, guiding context-aware deployment for neuro-oncology applications.