Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

MCA-RG: Enhancing LLMs with Medical Concept Alignment for Radiology Report Generation

Created by
  • Haebom

Author

Qilong Xing, Zikai Song, Youjia Zhang, Na Feng, Junqing Yu, Wei Yang

Outline

This paper proposes a Medical Concept Aligned Radiology Report Generation (MCA-RG) framework to address the challenges of accurately mapping pathological and anatomical features to textual descriptions in applying large-scale language models (LLMs) to medical image report generation (RRG) and the difficulty of generating accurate diagnostic reports due to semantically irrelevant feature extraction. MCA-RG leverages two curated concept banks—a pathology bank containing pathology-related knowledge and an anatomy bank containing anatomical descriptions—to explicitly align visual features with individual medical concepts. It proposes an anatomy-based contrastive learning procedure to improve anatomical feature generalization and a matching loss for pathological features to prioritize clinically relevant regions. Furthermore, it employs a feature gating mechanism to filter out low-quality concept features and utilizes visual features corresponding to individual medical concepts to guide the report generation process. Experiments on two publicly available benchmarks, MIMIC-CXR and CheXpert Plus, demonstrate that MCA-RG achieves excellent performance.

Takeaways, Limitations

•
Takeaways:
◦
Improved accuracy of medical image report generation: Improved accuracy of medical image report generation through accurate mapping of pathological and anatomical features to text descriptions.
◦
Solving the semantically irrelevant feature extraction problem: We solved the semantically irrelevant feature extraction problem by explicitly aligning medical concepts with visual features.
◦
Leveraging Concept-Based Knowledge: We leveraged a curated bank of medical concepts to enable knowledge-based report generation.
◦
Performance enhancement through contrastive learning and matching loss: We improved the generalization performance and clinical relevance of the model through anatomy-based contrastive learning and matching loss to pathological features.
◦
Low-quality feature filtering: Low-quality concept features were effectively removed through a feature gating mechanism.
◦
We demonstrate the effectiveness of the proposed method by achieving excellent performance on two public benchmarks: MIMIC-CXR and CheXpert Plus.
•
Limitations:
◦
Dependency of the curated concept bank: Performance may depend on the quality and scope of the curated concept bank. Incompleteness or bias in the bank can affect model performance.
◦
Adaptability to new medical concepts: Further research is needed on adaptability to new medical concepts or diseases.
◦
Interpretability: Further research is needed to increase the interpretability of the model's decision-making process.
◦
Clinical applicability: Further validation of applicability in real-world clinical settings is needed.
👍
Made with Slashpage