Utilizing Longitudinal Chest X-Rays and Reports to Pre-Fill Radiology Reports

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Radiology Reporting

Background:

• Speech recognition software reduces radiology reporting time but persistent communication errors remain a concern.
• Existing methods for generating medical reports lack the ability to utilize longitudinal patient data.
• Pre-filling radiology reports is a promising strategy to mitigate reporting errors.

Approach:

• A transformer-based model was developed to pre-fill the "findings" section of radiology reports using longitudinal multi-modal data.
• The model utilizes previous and current chest X-ray (CXR) images and prior visit reports from the MIMIC-CXR dataset.
• A novel dataset, Longitudinal-MIMIC, was created comprising data from 26,625 patients.

Key Points:

• The model employs a cross-attention multi-modal fusion module and a hierarchical memory-driven decoder.
• Unlike previous methods, this approach exploits longitudinal patient visit information.
• Experimental results demonstrate superior performance compared to recent state-of-the-art methods.

Conclusions:

• The proposed method effectively pre-fills radiology report findings by leveraging longitudinal multi-modal data.
• This approach addresses the gap in utilizing historical patient data for report generation.
• The developed model shows significant potential in enhancing the accuracy and efficiency of radiology reporting.