Learning from dispersed manual annotations with an optimized data weighting policy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models are crucial for medical image analysis but struggle with biased and imbalanced training data.
• Combining data from different sources can increase training data but introduces challenges like varying class balance and labeling protocols.
• Existing importance sampling methods do not adequately address imbalanced data from sources with distinct labeling.

Purpose of the Study:

• To introduce a novel sampling policy, Adaptive Stochastic Policy (ASP), for training deep learning models on imbalanced, multi-source medical imaging data.
• To adapt training strategies based on subject, data source, and dynamic criteria using reinforcement learning principles.
• To evaluate ASP's effectiveness in multiorgan abdominal computed tomography segmentation.

Main Methods:

• Developed Adaptive Stochastic Policy (ASP), a reinforcement learning-inspired sampling strategy.
• Applied ASP to multiorgan abdominal CT segmentation tasks.
• Conducted fivefold cross-validation on 840 subjects from 10 data sources and external validation on 20 subjects from one source.

Main Results:

• ASP achieved an average Dice score of 0.8261 in cross-validation, outperforming the Upper Confident Bound (UCB) baseline (0.8135).
• On withheld test datasets, ASP reached a mean Dice score of 0.8265, compared to UCB's 0.8077.
• Statistical tests confirmed the significant improvement provided by ASP over the baseline.

Conclusions:

• ASP offers a flexible and effective reweighting technique for training deep learning models.
• The proposed method successfully adapts sample importance, enhancing performance on complex segmentation tasks.
• ASP demonstrates its capability in handling multisite, multiorgan, and varying-size medical image segmentation challenges.