Extractive Radiology Reporting With Memory-Based Cross-Modal Representations.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Radiology report generation (RRG) is vital for medical diagnosis but current autoregressive models risk invalid content and slow processing.
• Advanced models like LLMs improve RRG but can hallucinate and remain slow.

Purpose of the Study:

• To develop a new extractive radiograph reporting (ERR) workflow to address limitations of generative RRG.
• To design a framework for efficient and accurate sentence extraction for radiology reports.

Main Methods:

• Proposed a novel extractive radiograph reporting (ERR) workflow.
• Developed a framework utilizing a memory module for storing medical information and enhancing cross-modal representations.
• Employed sentence matching for efficient extraction from existing radiological cases.

Main Results:

• The ERR approach demonstrated superior performance compared to strong baselines on two benchmark datasets.
• Achieved comparable results to state-of-the-art generative models in radiology report generation.
• Confirmed reliable content generation and high training/inference efficiency.

Conclusions:

• The proposed ERR workflow offers a reliable and efficient alternative to traditional generative RRG methods.
• ERR ensures accuracy and speed, mitigating risks of invalid content and hallucinations.
• This approach enhances the practical application of AI in medical reporting.