Related Experiment Video
Updated: Jun 30, 2026
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.0K
RobMedNAS: searching robust neural network architectures for medical image synthesis
Jinnian Zhang1, Weijie Chen1, Tanmayee Joshi1
1Electrical and Computer Engineering, University of Wisconsin-Madison, Madison, United States of America.
Biomedical Physics & Engineering Express
|August 13, 2024
Summary
This study enhances U-Net model robustness for medical image synthesis using RobMedNAS, a novel neural architecture search. RobMedNAS optimizes U-Net to resist adversarial attacks, ensuring accurate medical image processing.
Area of Science:
- Medical image processing
- Artificial intelligence in healthcare
- Deep learning for medical imaging
Background:
- U-Net models are crucial for medical image synthesis but vulnerable to adversarial perturbations.
- Ensuring robustness is vital for reliable clinical applications of synthesized medical images.
Purpose of the Study:
- To introduce RobMedNAS, a neural architecture search strategy, for developing robust U-Net models.
- To evaluate the resilience of RobMedNAS-optimized U-Net models against adversarial attacks in medical image synthesis.
Main Methods:
- Retrospective analysis of synthesized CT from MRI data.
- Evaluation using Dice coefficient and mean absolute error metrics.
- Comparison of traditional U-Net models with RobMedNAS-optimized models under adversarial conditions.
Main Results:
- RobMedNAS effectively enhances U-Net model resilience to adversarial perturbations.
- Optimized models maintain high accuracy in medical image synthesis.
- Demonstrated improvement in robustness without sacrificing performance.
Conclusions:
- RobMedNAS offers a viable strategy for creating robust U-Net architectures for medical image synthesis.
- This approach advances the reliability of AI-driven medical image processing.
- The findings pave the way for more secure and dependable medical imaging solutions.
