A Disease-Aware Dual-Stage Framework for Chest X-ray Report Generation

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Computer Vision and Natural Language Processing

Background:

• Automated radiology report generation from chest X-rays is crucial for efficiency.
• Current AI models struggle with disease-specific features and accurate clinical report generation.
• Existing methods lack sufficient disease-awareness and vision-language alignment.

Purpose of the Study:

• To develop a novel dual-stage, disease-aware framework for enhanced chest X-ray report generation.
• To improve the clinical accuracy and linguistic quality of AI-generated radiology reports.
• To address limitations in current AI models regarding disease representation and medical image analysis.

Main Methods:

• A dual-stage framework was proposed, learning Disease-Aware Semantic Tokens (DASTs) via cross-attention and multi-label classification.
• Vision and language representations were aligned using contrastive learning in Stage 1.
• A Disease-Visual Attention Fusion (DVAF) module and Dual-Modal Similarity Retrieval (DMSR) mechanism were introduced in Stage 2 for contextual guidance.

Main Results:

• The proposed disease-aware framework achieved state-of-the-art performance on benchmark datasets (CheXpert Plus, IU X-ray, MIMIC-CXR).
• Significant improvements in clinical accuracy and linguistic quality of generated reports were observed.
• The model demonstrated enhanced ability to identify and report critical pathological features.

Conclusions:

• The novel dual-stage disease-aware framework effectively addresses limitations in current chest X-ray report generation models.
• This approach offers a promising solution for improving AI-assisted radiology by enhancing report accuracy and clinical relevance.
• The framework shows potential to significantly reduce radiologists' workload and improve patient care through faster, more accurate diagnoses.