Encoder-decoder models for chest X-ray report generation perform no better than unconditioned baselines.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• High-quality radiology reports are crucial for patient diagnosis and care.
• Automated report generation can reduce radiologist workload and prevent errors.
• Machine learning (ML) models are being developed to automatically generate radiology reports from X-ray images.

Purpose of the Study:

• To evaluate the performance of ML models for chest X-ray radiology report generation.
• To compare encoder-decoder models like 'Show, Attend and Tell' (SA&T) against unconditioned baselines.

Main Methods:

• Utilized the IU chest X-ray dataset for evaluation.
• Compared SA&T and other encoder-decoder models with unconditioned models that do not use image input.
• Assessed performance using diagnostic accuracy, BLEU-4, and METEOR metrics.

Main Results:

• An unconditioned model achieved a diagnostic accuracy of 0.91, significantly outperforming SA&T (0.877) and other ML models (p < 0.001).
• The unconditioned model also showed superior performance on BLEU-4 and METEOR metrics compared to SA&T.
• An unconditioned version of SA&T achieved comparable diagnostic accuracy to the original SA&T (0.862 vs. 0.877, p ≥ 0.05).

Conclusions:

• Popular encoder-decoder ML models for radiology report generation may not outperform simpler unconditioned baselines.
• The findings suggest that current image-based ML models may not be effectively leveraging visual information for report generation.
• Further research is needed to develop ML models that genuinely benefit from image input for improved radiology reporting.