Policy Learning for Actively Labeled Sample Selection on Lumbar Semi-supervised Classification.

Area of Science:

• Medical Imaging
• Machine Learning
• Data Science

Background:

• Acquiring labeled medical data is challenging due to annotation costs.
• Semi-supervised learning leverages unlabeled data but is sensitive to labeled data quality.
• Active learning enhances model performance by selecting informative samples.

Purpose of the Study:

• To propose a unified semi-supervised active learning architecture (RL-based SSAL).
• To improve the efficiency and performance of medical image analysis models with limited labeled data.

Main Methods:

• Developed a reinforcement learning-based approach for active sample selection.
• Designed a novel reward function combining predictive confidence and uncertainty.
• Alternately trained a semi-supervised network and performed active sample selection.

Main Results:

• The RL-based SSAL achieved over 3% performance improvement compared to a semi-supervised baseline.
• The model demonstrated superiority over other active learning methods.
• Achieved 89.32% accuracy with only 200 labeled samples, comparable to using the full dataset.

Conclusions:

• The proposed RL-based SSAL effectively reduces the need for extensive labeled medical data.
• Reinforcement learning offers a powerful strategy for informative sample selection in semi-supervised learning.
• This method significantly advances the application of machine learning in medical diagnostics.