Image synthesis with disentangled attributes for chest X-ray nodule augmentation and detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Lung nodule detection in chest X-rays (CXRs) is crucial for early lung cancer screening.
• Deep learning-based Computer-Assisted Diagnosis (CAD) systems require large, diverse datasets for training robust models.
• Existing lung nodule synthesis methods struggle to generate realistic nodules with desired shape and size attributes.

Purpose of the Study:

• To develop a novel lung nodule synthesis framework for data augmentation in CXR analysis.
• To generate realistic lung nodules with controllable shape, size, and texture attributes.
• To improve the performance of lung nodule detection CAD systems.

Main Methods:

• A novel framework decomposes nodule attributes into shape, size, and texture.
• A Generative Adversarial Network (GAN)-based Shape Generator creates diverse nodule shape masks.
• Size Modulation allows pixel-level control of nodule diameters.
• A coarse-to-fine gated convolutional Texture Generator synthesizes visually plausible nodule textures.

Main Results:

• The proposed framework generates lung nodules with enhanced image quality, diversity, and controllability.
• Synthesized nodule CXR images effectively compensate for nodules missed in detection tasks.
• Experimental results demonstrate significant improvements in nodule detection performance using the data augmentation strategy.

Conclusions:

• The novel lung nodule synthesis framework offers a powerful tool for data augmentation in medical imaging.
• Controllable synthesis of realistic lung nodules can significantly enhance the accuracy of CAD systems.
• This approach holds promise for improving early lung cancer detection rates through improved AI models.