Knowledge Graph-Based Few-Shot Learning for Label of Medical Imaging Reports

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Data Science

Background:

• Manual labeling of large medical imaging datasets for AI applications is a significant bottleneck.
• Existing AI models struggle with the scale and complexity of automatic imaging report labeling.

Purpose of the Study:

• To develop an effective data augmentation strategy for few-shot learning in automatic radiology report labeling.
• To leverage knowledge graphs (KGs) to enhance data augmentation for AI models.

Main Methods:

• Constructed a KG for lumbar spine X-ray images and annotated 2000 data points.
• Augmented training data using KG attributes (synonyms/replacements) and trained a BERT model.
• Evaluated performance with varying augmentation ratios and methods, focusing on precision and F1 scores.

Main Results:

• Data augmentation significantly improved model performance, with precision reaching 0.932 and F1 score 0.881 at a 1:100 ratio.
• Combining synonym and replacement augmentation yielded the highest F1 score (0.881).
• Iterative data supplementation for underperforming nodes further boosted the average F1 score to 0.979.

Conclusions:

• Knowledge graph-driven data augmentation is effective for training AI models on limited datasets for radiology report labeling.
• This method substantially reduces manual labeling efforts and improves labeling accuracy and efficiency.
• Presents a viable strategy for advancing AI applications in medical image report automation.