An Organ-Aware Diagnosis Framework for Radiology Report Generation

Area of Science:

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

Background:

• Radiology report generation (RRG) is vital for radiologist efficiency.
• Current RRG methods often adapt image captioning without specific medical insights.
• There's a need for RRG systems that provide detailed, organ-specific diagnostic information.

Purpose of the Study:

• To develop an AI model for generating comprehensive radiology reports.
• To improve the efficiency and accuracy of RRG by incorporating organ-wise information.
• To create a system that differentiates between normal and abnormal findings for each organ.

Main Methods:

• Proposed Organ-aware Diagnosis (OaD) model for RRG.
• Utilized task distillation (TD) to extract organ-level descriptions from reports.
• Developed an organ-aware report generation module for specific organ descriptions and simulated clinical scenarios.
• Implemented an auto-balance mask loss for balanced training across organs and findings.

Main Results:

• OaD demonstrated superior performance compared to state-of-the-art methods.
• Achieved a 3.4% BLEU-1 improvement on the MIMIC-CXR dataset.
• Obtained a 2.0% BLEU-2 improvement on the IU-Xray dataset.
• Successfully generated organ-specific descriptions, enhancing report detail.

Conclusions:

• OaD offers a practical and effective approach to radiology report generation.
• The organ-aware strategy significantly improves RRG performance on benchmark datasets.
• This method has the potential to substantially increase radiologist efficiency and diagnostic precision.