Automated Radiology Report Generation: A Review of Recent Advances.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Medical imaging departments face increasing demands, impacting radiologist efficiency and report timeliness.
• Artificial intelligence offers potential solutions for automatic radiology report generation (ARRG).
• Research in ARRG has rapidly expanded due to technological advancements.

Purpose of the Study:

• To conduct a methodological review of contemporary automatic radiology report generation (ARRG) approaches.
• To assess datasets, deep learning methods, model architectures, clinical knowledge integration, and evaluation techniques in ARRG.
• To identify insights from top-performing models and highlight future research directions.

Main Methods:

• Systematic review of ARRG literature.
• Analysis of datasets based on availability, size, and adoption.
• Examination of deep learning training methods (e.g., contrastive, reinforcement learning).
• Exploration of model architectures (CNN, transformer variations).
• Review of techniques for integrating clinical knowledge (multimodal inputs, knowledge graphs).
• Scrutiny of evaluation techniques (NLP metrics, clinical reviews).

Main Results:

• Analysis of quantitative results from reviewed ARRG models.
• Identification of top-performing models and their characteristics.
• Insights derived from the performance of state-of-the-art ARRG systems.

Conclusions:

• Current ARRG approaches vary in datasets, training methods, architectures, and evaluation.
• Further development is needed in utilizing diverse datasets and refining evaluation metrics.
• Future directions include expanding datasets to include more radiological modalities and enhancing evaluation methodologies for ARRG.