TRINet: Team Role Interaction Network for automatic radiology report generation.

Area of Science:

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

Background:

• Automatic radiology report generation is crucial for supporting expert diagnosis.
• Current models struggle with uncertainty and report stability due to complex medical image interpretation.

Purpose of the Study:

• To develop a robust AI model for automatic radiology report generation.
• To address aleatoric and epistemic uncertainty in current generative models.

Main Methods:

• Proposed the Team Role Interaction Network (TRINet) with multiple Team-member and one Team-leader role models.
• Introduced a Cross-Modal Communication Mechanism (CMCM) for information exchange between Team-members.
• Utilized a Multi-Modal Fusion Mechanism (MMFM) for aggregating knowledge and generating final reports.

Main Results:

• TRINet demonstrated state-of-the-art performance on IU X-ray and MIMIC-CXR datasets.
• Achieved a BLEU-4 score of 0.144 on MIMIC-CXR, surpassing previous methods by 3.6 points.
• Showcased improved accuracy and robustness by leveraging complementary information between modalities.

Conclusions:

• TRINet effectively simulates collaborative expert diagnosis for improved radiology report generation.
• The model enhances system accuracy and robustness by integrating diverse information sources.
• TRINet offers a promising solution for stable and reliable AI-assisted radiological reporting.