New self-supervised method enriches medical imaging reports by adding omitted positive findings, boosting vision-language model performance by up to 7.47%

SemEnrich addresses the bias in medical datasets where clinicians predominantly report abnormalities while omitting positive/neutral findings

The method uses semantic clustering of report sentences to automatically enrich training data with relevant observations from different clusters

Testing showed significant improvements: 5.63% gain on COMET score, 7.47% on RadGraph-F1, 7.40% on Sentence BLEU, and 5.30% on CheXbert-F1

Ablation studies confirmed that semantic clustering drives improvements, not random data augmentation

Researchers also developed a way to incorporate semantic cluster information into reward design for GRPO training