AI-Assisted Pneumonia Detection, Localisation and Report Generation from Chest X-rays

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Machine Learning for Healthcare

Background:

• Pneumonia detection in chest X-rays (CXRs) faces challenges due to inter-observer variability and overlapping radiographic patterns.
• Deep learning (DL) solutions show promise but are limited by generalisability and explainability, hindering clinical adoption.
• Current methods often rely on rule-based natural language processing (rNLP) for report analysis, which can be suboptimal.

Purpose of the Study:

• To develop and evaluate a holistic deep learning (DL)-based computer-aided diagnosis (CAD) pipeline for pneumonia detection, localisation, and structured report generation from CXRs.
• To assess the impact of large language model (LLM)-driven relabelling on diagnostic sensitivity compared to traditional rNLP labels.
• To improve the generalisability and explainability of DL models for pneumonia detection in CXRs.

Main Methods:

• Curated a large composite dataset of 922,634 publicly available CXRs.
• Relabelling MIMIC-CXR radiology reports using a local LLM to generate pneumonia labels.
• Trained DenseNet-121 classifiers on different data configurations (MIMIC-CXR with rNLP and LLM labels, supplemented with VinDr-CXR data).
• Utilized Gradient-weighted Class Activation Mapping (Grad-CAM) for visual explainability and lung zone-based localisation.

Main Results:

• LLM-driven relabelling significantly improved human-label agreement (96.5% vs 72.5%, P=1.66×10 ^-11).
• The best-performing model (MIMIC-CXR (LLM) + VinDr-CXR) achieved 82.08% sensitivity and 81.97% precision, outperforming radiologist sensitivity ranges and CheXNet.
• Grad-CAM localisation achieved a moderate F1-score of 52.9%, indicating alignment with pathological regions.

Conclusions:

• LLM-driven label curation combined with DL surpasses conventional rNLP and radiologist performance in pneumonia detection.
• The developed CAD pipeline demonstrates potential for rapid triage, automated report drafting, and real-time pneumonia surveillance.
• This approach advances high-quality data integration in predictive medical imaging, streamlining radiology workflows and mitigating diagnostic errors.