Abnormality-Aware Knowledge Graph and Hierarchical Cross-Attention in LLM for Radiology Report Generation

Radiologist workload: Automatic radiology report generation is motivated by reducing radiologists’ workload while maintaining clinical accuracy.

Normal bias in medical datasets: Existing radiology report generation methods are limited by normal bias in medical datasets.

Static multi-scale feature utilization: Existing methods are limited by static use of multi-scale visual features.

Radiologist workload: Automatic radiology report generation is motivated by reducing radiologists’ workload while maintaining clinical accuracy.

Normal bias in medical datasets: Existing radiology report generation methods are limited by normal bias in medical datasets.

Abnormality-aware knowledge graph: The paper proposes constructing an abnormality-aware knowledge graph that fuses image features and mitigates normal bias.

Hierarchical cross-attention LLM decoder: The paper proposes a hierarchical cross-attention mechanism in the LLM decoder to dynamically select global and regional visual features during word generation.

Abnormality-aware knowledge graph: The paper proposes constructing an abnormality-aware knowledge graph that fuses image features and mitigates normal bias.

Hierarchical cross-attention LLM decoder: The paper proposes a hierarchical cross-attention mechanism in the LLM decoder to dynamically select global and regional visual features during word generation.

State-of-the-art performance:

Clinical efficacy F1 improvement:

Enhanced clinical reliability: