Joint Imbalance Adaptation for Radiology Report Generation.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Natural Language Processing for Healthcare
• Machine Learning for Clinical Decision Support

Background:

• Radiology report generation faces significant data imbalance challenges, with infrequent medical tokens and a prevalence of normal findings.
• Existing methods often overlook these specific imbalance issues, potentially leading to underperformance in generating accurate clinical reports.
• Addressing data imbalance is crucial for improving the reliability and clinical utility of automated radiology report generation systems.

Purpose of the Study:

• To propose a novel model, Joint Imbalance Adaptation (JIMA), to tackle both token and label imbalance in radiology report generation.
• To enhance the robustness of report generation models by mitigating overfitting to frequent data patterns and underfitting to infrequent ones.
• To improve the generation of precise and clinically relevant descriptions from radiological images.

Main Methods:

• Developed the Joint Imbalance Adaptation (JIMA) model, specifically designed to leverage token and label imbalance factors.
• Implemented a hard-to-easy learning strategy to guide the model's focus towards less frequent labels and clinical tokens.
• Evaluated JIMA's performance on the IU X-ray and MIMIC-CXR datasets using standard evaluation metrics.

Main Results:

• JIMA demonstrated significant average improvements ranging from 16.75% to 50.50% across various evaluation metrics on benchmark datasets.
• Ablation studies and human evaluations confirmed that improvements stem from enhanced performance on infrequent tokens and abnormal radiological entries.
• The model successfully generated more clinically accurate radiology reports, particularly for underrepresented data patterns.

Conclusions:

• The Joint Imbalance Adaptation (JIMA) model effectively addresses data imbalance in radiology report generation.
• The proposed imbalance learning strategy offers a promising direction for improving model performance on infrequent data, leading to more accurate clinical reports.
• JIMA's success highlights the importance of tailored approaches to handle data heterogeneity in medical AI applications.