Related Experiment Video
Updated: Jan 13, 2026
Author Spotlight: An Efficient and Robust Software for Automated Fusion of Multiple Preclinical Imaging Modalities
Published on: October 27, 2023
SupReMix: Supervised contrastive learning for medical imaging regression with mixup
Yilei Wu1, Zijian Dong2, Chongyao Chen3
1Centre for Sleep and Cognition & Centre for Translational Magnetic Resonance Research, Yong Loo Lin School of Medicine, National University of Singapore, Singapore; Healthy Longevity & Human Potential Translational Research Program and Department of Medicine, Yong Loo Lin School of Medicine, National University of Singapore, Singapore.
We introduce SupReMix, a novel contrastive learning method for medical image regression. SupReMix improves feature representation by incorporating ordinality and hardness, leading to significantly better diagnostic predictions.
Area of Science:
- Medical Image Analysis
- Computer-Aided Diagnosis
- Deep Learning
Background:
- Regression is vital for quantitative medical image analysis (e.g., age prediction, cardiac function).
- Current deep learning methods often overlook feature representation quality, hindering robust clinical predictions.
- Representation learning for classification doesn't directly translate well to regression tasks.
Purpose of the Study:
- To enhance deep learning-based medical image regression by addressing limitations in feature representation.
- To propose a novel contrastive learning framework that accounts for ordinality and hardness in regression tasks.
- To improve the accuracy and robustness of computer-aided diagnosis through better feature learning.
Main Methods:
- Proposed Supervised Contrastive Learning for Medical Imaging Regression with Mixup (SupReMix).
- SupReMix utilizes anchor-inclusive mixtures as hard negative pairs and anchor-exclusive mixtures as hard positive pairs.
- Integrates richer ordinal information and hardness into contrastive learning at the embedding level.
Main Results:
- SupReMix fosters continuous ordered representations in the latent space.
- Demonstrated significant improvements in regression performance across six diverse medical imaging datasets (MRI, X-ray, ultrasound, PET).
- Outperformed existing methods by optimizing feature representation quality for regression tasks.
Conclusions:
- The proposed SupReMix method effectively enhances medical image regression by learning superior feature representations.
- Addressing ordinality and hardness in contrastive learning is crucial for advancing medical image regression.
- SupReMix offers a promising direction for improving computer-aided diagnosis through deep representation learning.
