OS-RRG: Observation State-Aware Radiology Report Generation With Balanced Diagnosis and Attention Intervention.

Area of Science:

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

Background:

• Automated radiology report generation (RRG) aims to assist radiologists by producing detailed clinical radiography descriptions and diagnoses.
• Current RRG methods face challenges in accurately identifying observation states (OSs) and aligning them with textual descriptions, often due to data imbalances.

Purpose of the Study:

• To introduce a novel framework, observation state-aware radiology report generator (OS-RRG), to improve both the identification and alignment of observation states in RRG.
• To address interclass and intraclass imbalances in observation states that hinder performance in existing methods.

Main Methods:

• Developed a state-aware balancing diagnosis (SBD) module to mitigate interclass and intraclass imbalances in observation states.
• Implemented a state-guided attention intervention (SAI) technique for dynamic focus adjustment on critical diagnostic features.
• Employed a task-specific learning paradigm to decouple identification and alignment processes into independent pathways.

Main Results:

• The OS-RRG framework demonstrated superior diagnostic accuracy compared to existing state-of-the-art techniques on the MIMIC-CXR and IU-Xray benchmarks.
• The proposed SBD module and SAI technique effectively addressed observation state imbalances and improved feature focus.
• The decoupled learning paradigm significantly enhanced overall RRG performance.

Conclusions:

• The OS-RRG framework offers a significant advancement in automated radiology report generation by improving observation state identification and alignment.
• The novel techniques presented can enhance the clinical accuracy and reliability of AI-generated radiology reports.
• This work provides a foundation for more robust and accurate RRG systems, ultimately supporting clinical decision-making.