Robustness Metrics for Tumor Volume Delineation CNN

Research Internship at Karlsruhe Institute of Technology (KIT) Advisor: Prof. Martin Frank

Automated tumor delineation is critical for radiation therapy, but deep learning models must be robust to noise and artifacts. This project stress-tested the state-of-the-art nnU-Net architecture.

Key Contributions:

• Designed robustness metrics for automated head and neck tumor volume delineation.
• Augmented Computer Tomography (CT) data using MONAI (Medical Open Network for AI).
• Ran inference pipelines using PyTorch to evaluate model failure modes.
• Found that standard models were not robust to significant augmentations, highlighting the need for uncertainty quantification in medical AI.