Semantically Redundant Training Data Removal and Deep Model Classification Performance: A Study with Chest X-rays.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Machine Learning

Background:

• Deep learning (DL) models excel at learning from complex data, with performance often linked to data volume.
• Semantic redundancy, or repetitive information in training data, can hinder DL model performance and generalizability.
• In medical imaging, semantic redundancy may arise from similar disease presentations and be exacerbated by data augmentation.

Approach:

• Proposed an entropy-based sample scoring method to identify and remove semantically redundant data from training sets.
• Utilized the NIH chest X-ray dataset for demonstrating the effectiveness of the proposed approach.
• Compared model performance trained on the full dataset versus a curated subset of informative samples.

Key Points:

• Models trained on the informative subset showed significantly improved recall during internal testing (0.7164 vs 0.6597).
• External testing also demonstrated superior performance for the model trained on the informative subset (0.3185 vs 0.2589).
• The findings highlight the critical role of information-oriented sample selection over simply using all available data.

Conclusions:

• Data selection strategies are crucial for optimizing deep learning model performance in medical imaging.
• Reducing semantic redundancy through targeted sample selection can lead to more generalizable and accurate models.
• This approach offers a valuable alternative to conventional methods that rely on maximizing training data volume.