GANDALF: Graph-based transformer and Data Augmentation Active Learning Framework with interpretable features for...

Area of Science:

• Machine Learning
• Medical Imaging
• Computer-Aided Diagnosis

Background:

• Active learning (AL) reduces annotation costs by selecting informative samples for training machine learning models.
• Data augmentation further expands datasets in low-data scenarios.
• Combining these techniques is intuitive for improving AL system performance.

Purpose of the Study:

• To propose GANDALF (Graph-based TrANsformer and Data Augmentation Active Learning Framework), a novel approach for multi-label active learning.
• To address limitations of conventional AL methods in multi-label settings, particularly for medical images where samples can have multiple disease labels.
• To improve the performance, learning rates, and robustness of computer-aided diagnosis systems.

Main Methods:

• Representing disease labels as graph nodes and utilizing graph attention transformers (GAT) to learn inter-label relationships.
• Identifying informative samples by aggregating GAT representations.
• Generating augmented samples from a learned latent space and selecting informative ones using a novel multi-label informativeness score that ensures non-redundancy and training contribution.

Main Results:

• GANDALF demonstrated improved performance over state-of-the-art multi-label AL techniques on chest X-ray and MedMNIST datasets.
• The method showed enhanced learning rates and robustness in diagnostic tasks.
• The novel informativeness score effectively identified valuable augmented samples.

Conclusions:

• GANDALF effectively integrates informative sample selection and data augmentation in a multi-label AL framework.
• The approach significantly improves the performance of computer-aided diagnosis systems with limited labeled data.
• GANDALF offers a robust solution for multi-label learning in medical image analysis.