Semi-Supervised Learning in Medical Images Through Graph-Embedded Random Forest.

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Computer Vision

Background:

• Medical imaging analysis faces challenges due to the scarcity of labeled data, requiring specialized expertise.
• This data limitation is particularly critical in brain image analysis, including retinal vasculature assessment, a key indicator of Central Nervous System (CNS) vascular health.

Purpose of the Study:

• To develop a novel semi-supervised learning algorithm to enhance random forest performance with limited labeled data.
• To address the bottleneck in random forest's information gain calculation for improved reliability in low-data scenarios.

Main Methods:

• The proposed algorithm exploits the local structure of unlabeled data to boost performance.
• It replaces the standard information gain calculation with a graph-embedded entropy measure.
• The training process of the standard random forest is modified to integrate this new approach.

Main Results:

• The novel algorithm significantly improves performance in medical imaging analysis and machine learning benchmarks.
• It demonstrates superior results compared to standard methods when labeled data is insufficient.
• The method maintains the computational efficiency and robustness against overfitting inherent to random forests.

Conclusions:

• The developed semi-supervised learning algorithm effectively overcomes the challenge of limited labeled data in medical imaging.
• Graph-embedded entropy offers a more reliable measure for information gain in low-data regimes.
• This approach enhances the utility of random forests for complex tasks like brain image analysis.