Active multitask learning with uncertainty-weighted loss for coronary calcium scoring.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Cardiovascular Disease Risk Assessment
• Radiology Workflow Optimization

Background:

• Coronary artery calcification (CAC) scoring is crucial for cardiovascular event risk prediction.
• Automated CAC quantification can enhance radiologist efficiency but requires large, expert-annotated datasets.
• Active learning and multitask learning offer potential solutions to reduce data requirements and improve model performance.

Purpose of the Study:

• To develop an uncertainty-weighted multitask learning model for automated CAC scoring.
• To simultaneously train segmentation tasks for coronary artery regions and calcifications.
• To reduce the number of required training samples and labeling time through an active learning strategy.

Main Methods:

• Proposed an uncertainty-weighted multitask learning model for ECG-gated cardiac CT.
• Trained the model on coronary artery region segmentation (weak labels) and CAC segmentation (strong labels) concurrently.
• Compared the model against single-task and sequential-task U-Net models on three independent datasets.
• Analyzed the impact of image quality factors on model performance.

Main Results:

• Joint learning of segmentation tasks enhanced CAC scoring performance.
• The model achieved optimal performance using only 12% of the training data and one-third of the labeling time in an active learning setting.
• Image noise, metal artifacts, and anatomical abnormalities were identified as key factors affecting model performance.

Conclusions:

• The proposed uncertainty-weighted multitask learning approach effectively improves CAC scoring.
• This method leverages shared features from joint segmentation tasks for better performance.
• The active learning scenario significantly reduces labeling costs and data requirements for CAC analysis.