A disentangled generative model for disease decomposition in chest X-rays via normal image synthesis

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Interpreting medical images like chest X-rays (CXRs) is complex due to anatomical superposition, low resolution, and poor boundaries.
• Existing computer-aided diagnosis tools primarily focus on disease classification, not on disentangling normal anatomy from pathology.

Purpose of the Study:

• To develop a novel deep disentangled generative model (DGM) for simultaneous generation of disease residue maps and patient-specific normal CXR images from abnormal inputs.
• To enhance the interpretability and efficiency of CXR analysis in clinical practice.

Main Methods:

• Proposed a deep disentangled generative model (DGM) with three encoder-decoder branches: normal CXR synthesis (adversarial learning), disease residue map generation, and a robustness-enhancing branch.
• Employed self-reconstruction loss to ensure generated normal CXRs maintain visual structural similarity to original images.
• Evaluated the model on a large-scale chest X-ray dataset.

Main Results:

• The DGM successfully generated accurate disease residue/saliency maps, consistent with radiologist annotations.
• Produced radiorealistic, patient-specific normal CXR images.
• Demonstrated quantitative improvements in diagnostic performance for CXR classification and lung opacity detection.

Conclusions:

• The DGM effectively disentangles pathological information from normal anatomy in CXRs.
• Generated disease maps aid radiologists in improving reading efficiency.
• Synthesized normal CXRs support data augmentation and personalized disease studies.
• The model enhances diagnostic performance across various clinical applications.