IHRAS: Automated Medical Report Generation from Chest X-Rays via Classification, Segmentation, and LLMs

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Clinical Informatics

Background:

• Increasing demand for efficient and accurate Chest X-Ray (CXR) interpretation.
• Need for AI solutions to reduce radiologist workload and diagnostic variability.
• Lack of standardized, interpretable AI tools in radiological workflows.

Purpose of the Study:

• To introduce the Intelligent Humanized Radiology Analysis System (IHRAS), an AI framework for automated CXR analysis and report generation.
• To evaluate the diagnostic performance and report quality of IHRAS.
• To demonstrate a transparent and scalable AI solution for supporting radiological workflows.

Main Methods:

• Development of IHRAS, a modular framework integrating deep convolutional neural networks for classification, Grad-CAM for visualization, SAR-Net for segmentation, and a large language model (DeepSeek-R1) for report generation.
• Utilized the CRISPE prompt engineering framework for report generation with SNOMED CT terminology.
• Evaluated on the NIH ChestX-ray dataset, assessing diagnostic performance and report quality metrics (faithfulness, relevancy, alignment).

Main Results:

• IHRAS demonstrated consistent diagnostic performance across diverse demographic and clinical subgroups.
• The system generated high-fidelity, clinically relevant radiological reports.
• Report quality scores for faithfulness, relevancy, and alignment were strong.

Conclusions:

• IHRAS provides an automated, end-to-end solution for CXR analysis and report generation.
• The system offers a transparent and scalable approach to support radiological workflows.
• Highlights the importance of interpretability and standardization in clinical AI applications.