Adaptive Contrast for Image Regression in Computer-Aided Disease Assessment.

Area of Science:

• Medical imaging analysis
• Machine learning
• Computer-aided diagnosis

Background:

• Deep regression methods commonly use single loss functions (MSE, L1) for medical image analysis.
• Accurate bone mineral density (BMD) estimation and left-ventricular ejection fraction (LVEF) prediction are crucial for disease assessment.

Purpose of the Study:

• To introduce AdaCon, the first contrastive learning framework for deep image regression in medical applications.
• To improve the performance of regression tasks by incorporating label distance information into feature representations.

Main Methods:

• Developed a novel adaptive-margin contrastive loss for feature learning.
• Integrated a feature learning branch with a regression prediction branch.
• Utilized a plug-and-play module to enhance existing regression methods.

Main Results:

• AdaCon demonstrated superior performance in BMD estimation from X-ray images.
• Achieved a 3.3% relative improvement in Mean Absolute Error (MAE) for BMD estimation.
• Showcased a 5.9% relative improvement in MAE for LVEF prediction from echocardiogram videos.

Conclusions:

• AdaCon effectively enhances deep image regression for medical tasks.
• The framework offers a significant improvement over state-of-the-art methods in BMD and LVEF prediction.
• AdaCon's adaptive-margin contrastive loss provides a novel approach to incorporating label relationships in feature learning.