Pseudo-labeling and knowledge-guided contrastive learning for radiology report generation.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Natural Language Generation for Healthcare
• Deep Learning for Radiology

Background:

• Radiology report generation (RRG) aims to improve consistency and comprehensiveness in diagnostic interpretation.
• Current graph-based methods for RRG face limitations in semantic feature-text alignment, noisy annotations, and anatomical constraints.
• Addressing these limitations is crucial for advancing automated medical reporting.

Purpose of the Study:

• To propose a pseudo-label and knowledge-guided comparative learning (PKCL) framework for radiology report generation.
• To overcome semantic separation, annotation noise, and lack of anatomical constraints in existing RRG methods.
• To enhance the clinical accuracy and semantic consistency of generated radiology reports.

Main Methods:

• Developed a PKCL framework integrating dynamic query learning and knowledge-guided contrastive learning.
• Employed a trainable cross-modal query matrix (QM) for shared representations via self-attention mechanisms.
• Utilized pseudo-labels, adaptive feature fusion, and knowledge graphs (XRayVision) for improved learning and anatomical constraints.

Main Results:

• PKCL achieved state-of-the-art performance on IU-Xray and MIMIC-CXR datasets for both generation and clinical metrics.
• Outperformed previous methods like R2GEN and CMCL, achieving higher BLEU-1 and RL scores.
• Demonstrated robust generalization on out-of-domain datasets, even under low-resource conditions.

Conclusions:

• The PKCL framework significantly advances radiology report generation by maintaining semantic consistency.
• It effectively captures subtle relationships between radiological findings and textual descriptions.
• Represents a substantial improvement over existing methods for clinically relevant report generation.