Improving Radiology Report Generation Quality and Diversity through Reinforcement Learning and Text Augmentation

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Traditional radiology report generation (RRG) methods using vision encoder-decoder (VED) frameworks face challenges with report diversity and generalization.
• Existing benchmarks often fall short in capturing the nuances of comprehensive medical reporting.

Purpose of the Study:

• To improve the quality, diversity, and generalization capabilities of deep learning-based RRG.
• To establish new benchmarks for RRG accuracy and variability.

Main Methods:

• Implementation of reinforcement learning and innovative text augmentation techniques within a VED framework.
• Utilizing RadGraph as a key reward metric for training and evaluation.
• Benchmarking against established metrics like BLEU4, ROUGE-L, and F1CheXbert.

Main Results:

• The proposed VED model significantly surpasses existing benchmarks on MIMIC-CXR and Open-i datasets.
• Achieved F1-scores of 66.2 (CheXbert) and 37.8 (RadGraph) on MIMIC-CXR.
• Attained F1-scores of 54.7 (CheXbert) and 45.6 (RadGraph) on Open-i.

Conclusions:

• The integration of reinforcement learning and text augmentation represents a significant breakthrough in RRG.
• The enhanced models improve diagnostic precision and radiological interpretation in clinical settings.
• The research findings and code are publicly available to foster further advancements.