Research Poster
Project Overview
This project studies how to improve the robustness and generalization of 3D medical image segmentation. The team fine-tuned SegVol and compared two approaches: Sharpness-Aware Minimization (SAM) for optimization and Balanced Loss for severe foreground-background class imbalance.
Medical images often contain small foreground structures, such as organs or lesions, while most voxels belong to the background. Standard training can therefore under-segment small structures and converge to sharp minima that generalize poorly.
My Contribution
Sharpness-Aware Minimization
I was responsible for the Sharpness-Aware Minimization portion of the project.
- Applied SAM during SegVol fine-tuning in place of standard optimization.
- Used SAM to seek flatter minima in the loss landscape and improve model generalization.
- Evaluated segmentation performance using average Dice score across validation cases.
- Analyzed and presented the SAM method and experimental findings in the research poster.
Results
| Method | Average Dice Score |
|---|---|
| Baseline | 0.677 |
| Sharpness-Aware Minimization | 0.711 |
| Balanced Loss | 0.739 |
SAM improved the average Dice score from 0.677 to 0.711, showing that optimization toward flatter minima can improve segmentation generalization and robustness. Balanced Loss achieved the highest score by directly addressing class imbalance.
Research Team
Liangqiao Gui, Din Barrameda, Zihao Li, and Xin Wang, University at Albany.