*Background & Motivation*
Deep learning (DL) methods for medical image segmentation hold great potential for clinical applications but face significant challenges in real-world deployment. Two key limitations are their sensitivity to out-of-distribution data and the lack of interpretability, particularly regarding uncertainty quantification (UQ). UQ methods provide confidence estimates for model predictions, but their reliability under realistic data variations remains poorly understood. Ensuring that both accuracy and uncertainty estimates remain robust to distribution shifts is essential for developing trustworthy AI models in medical imaging.
*Objectives & Research Questions*
This project aims to systematically evaluate how common distortions in abdominal MRI scans impact both segmentation accuracy and uncertainty estimates in U-Net-based models. Specifically, we will:
- Simulate natural image alterations on a kidney MRI dataset to assess their effect on model performance.
- Develop and validate a novel metric to quantify the reliability of predictive uncertainty under data variations.
- Investigate how different training loss functions, along with Monte Carlo dropout and test-time augmentation, influence segmentation robustness.
- Explore potential strategies for improving the stability of uncertainty estimates in deep learning-based segmentation models.
Our central research question is: How can we design more robust deep learning models that maintain both high segmentation accuracy and reliable uncertainty estimates under real-world data variations?
*Impact & Expected Outcomes*
This work will contribute to the development of more interpretable and clinically reliable AI-driven segmentation systems. By systematically evaluating the robustness of UQ methods, we aim to inform best practices for integrating uncertainty-aware models into clinical workflows. Furthermore, the findings will provide insights into optimizing deep learning models for medical imaging applications, supporting broader adoption in healthcare settings.
We seek computational resources to accelerate this research, ensuring rigorous experimentation and high-quality analysis. The results will be shared through peer-reviewed publications and open-source contributions to foster collaboration in the medical AI community.