Advancement in medical report generation: current practices, challenges, and future directions

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology Reporting

Background:

• Radiologists' expertise is crucial for medical image analysis and disease diagnosis.
• Manual report generation is time-consuming and costly, hindering timely patient care.
• Automatic report generation streamlines the process and reduces the burden on radiologists.

Purpose of the Study:

• To systematically review and evaluate existing research on medical report generation methods.
• To analyze the performance and limitations of various deep learning models in this domain.
• To guide radiologists towards efficient and accurate diagnostic tools.

Main Methods:

• Systematic literature review following a defined protocol.
• Analysis and evaluation of 80+ research articles on medical report generation.
• Categorization and comparison of deep learning models including encoder-decoder, Transformers, attention mechanisms, RNN-LSTM, LLMs, and graph-based methods.

Main Results:

• Encoder-decoder frameworks are most common (45 articles) with 92-95% accuracy.
• Transformers-based models (20 articles) achieve ~91% accuracy.
• Other methods like attention mechanisms, RNN-LSTM, LLMs, and graph-based approaches show promising results but have limitations.

Conclusions:

• Deep learning models offer significant advancements in automatic medical report generation.
• Overfitting, bias, and high data dependency are key challenges impacting current methods.
• Future research should focus on addressing these limitations to enhance accuracy and efficiency in radiological diagnostics.