Mass Image Synthesis in Mammogram with Contextual Information Based on GANs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Scarcity of labeled lesion data limits deep learning in medical imaging.
• Synthesizing realistic breast mass images in mammograms is difficult due to variable patterns.

Purpose of the Study:

• To generate diverse, labeled medical images using contextual information in mammograms.
• To address the challenge of creating synthetic mass images with accurate shape, margin, and context.

Main Methods:

• Proposed a novel Generative Adversarial Network (GAN) approach for mass image synthesis.
• Implemented contextual infilling by inserting synthetic lesions into healthy mammograms.
• Incorporated realistic mass image features and masks into adversarial learning for improved generation.

Main Results:

• Demonstrated effectiveness on DDSM and a private hospital database.
• Achieved a 5.03% improvement in detection rate through data augmentation with synthetic images.
• Validated the approach through qualitative and quantitative evaluations.

Conclusions:

• The proposed method enhances dataset diversity for medical imaging.
• This approach is a significant step towards generating labeled breast mass images for precise detection.
• The method has potential for application in other medical imaging domains.