Related Experiment Video
Updated: Aug 5, 2026
06:48
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
Harmonization and Targeted Feature Dropout for Generalized Segmentation: Application to Multi-site Traumatic Brain
Yilin Liu1, Gregory R Kirk1, Brendon M Nacewicz1
1Waisman Laboratory for Brain Imaging and Behavior, University of Wisconsin-Madison, Madison, WI, USA.
Summary
This study introduces Targeted Feature Dropout (TFD) to improve the generalizability of deep learning models in medical imaging. TFD enhances model robustness for multi-site neuroimaging data, even with limited labeled training examples.
Area of Science:
- Medical Imaging
- Machine Learning
- Neuroimaging
Background:
- Deep learning models in medical imaging often lack generalizability across different data acquisition sites and protocols.
- Limited availability of expert-annotated medical images hinders the training of site-specific models.
- Multi-site data collection is crucial for statistical power but poses challenges for model transferability.
Purpose of the Study:
- To address the generalizability bottleneck in medical imaging by harmonizing multi-site data.
- To enhance model robustness against variations in target medical images using a novel technique.
- To improve the utilization of limited labeled source data for training robust models.
Main Methods:
- Harmonization of target medical imaging data using adversarial learning, specifically Cycle-consistent adversarial networks.
- Proposal of Targeted Feature Dropout (TFD), a technique guided by attention to stochastically remove discriminative features.
- Integration of TFD with existing models without increasing parameters or computational costs, suitable for small datasets.
Main Results:
- Demonstrated the feasibility of Cycle-consistent adversarial networks for harmonizing multi-site Magnetic Resonance Imaging (MRI) data.
- Showcased that TFD significantly improved the generalization of segmentation models on Traumatic Brain Injury (TBI) data.
- Achieved accuracy comparable to supervised learning methods despite using minimal labeled source data.
Conclusions:
- Adversarial learning and TFD are effective strategies for improving the generalizability of deep learning models in multi-site medical imaging studies.
- TFD offers a parameter-efficient method to enhance model robustness by leveraging attention mechanisms and dropout.
- The proposed approach shows promise for neuroimaging analysis, particularly in scenarios with limited labeled data, such as TBI datasets.
