Label correlated contrastive learning for medical report generation

Xinyao Liu, Junchang Xin, Bingtian Dai, Qi Shen, Zhihong Huang, Zhiqiong Wang

Abstract: Highlights•A label correlated contrastive learning approach is proposed to improve the model’s ability to distinguish the “hard” negative samples.•We construct correlations between multiple diseases to mine the disease information in an individual at a fine-grained level to improve the quality of the generated medical report.•Our proposed LACCOL learns richer semantic representations of abnormalities, addressing the problem of missing anomaly descriptions in reports.