Graph Node Based Interpretability Guided Sample Selection for Active Learning.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Supervised learning excels in medical image analysis but requires extensive labeled data.
• Active Learning (AL) optimizes performance with limited labels by selecting informative samples.
• Existing AL methods are primarily for single-label tasks, underperforming in multi-label scenarios like chest X-rays.

Purpose of the Study:

• To develop a novel sample selection approach for multi-label medical image analysis using graph analysis.
• To enhance the efficiency and accuracy of computer-aided diagnosis systems in resource-constrained settings.
• To address the limitations of conventional AL methods in handling multiple disease labels per sample.

Main Methods:

• Proposed a novel sample selection approach based on graph analysis for multi-label settings.
• Represented each class label as a node in a graph, with edge interactions reflecting model encoding similarities.
• Explored various graph aggregation techniques to identify the most informative samples for AL.
• Applied the method to public chest X-ray and medical image datasets.

Main Results:

• The proposed graph-based AL method demonstrated improved model performance compared to state-of-the-art techniques.
• Achieved better learning rates, indicating faster and more efficient model training.
• Showcased enhanced robustness in identifying informative samples across different medical imaging datasets.

Conclusions:

• The novel graph analysis approach effectively identifies informative samples in multi-label medical image analysis.
• This method offers a significant advancement over traditional AL techniques for complex diagnostic tasks.
• The findings suggest a more cost-effective and efficient pathway for developing high-performing AI in medical imaging.