Craniocaudal Mammograms Generation Using Image-to-Image Translation Techniques.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Breast cancer is a leading cause of death in women globally, necessitating improved early detection and prevention strategies.
• Access to large, reliable mammogram datasets for research is hindered by privacy concerns and data imbalances.
• Current public datasets for mammography are often unreliable and unbalanced, limiting AI model development.

Purpose of the Study:

• To develop a novel workflow for generating high-resolution synthetic mammograms.
• To enable precise control over synthetic breast features, including normal and tumor cases.
• To create a tool for augmenting or balancing existing datasets for machine learning applications.

Main Methods:

• Utilized a statistical generative model based on generative adversarial networks (GANs).
• Employed a 2D parametric model of the compressed breast in craniocaudal projection.
• Incorporated image-to-image translation techniques for feature control and generation.

Main Results:

• Generated synthetic mammograms accurately replicated features of real mammograms.
• Statistical analysis showed good correspondence between real and synthetic image statistics.
• Medical experts found synthetic and real mammograms statistically indistinguishable in most cases.

Conclusions:

• The proposed workflow successfully generates realistic synthetic mammograms with controllable features.
• Synthetic mammograms can enhance or balance datasets, facilitating the training of AI algorithms.
• These AI tools can assist radiologists in improving breast cancer classification and segmentation, enhancing diagnostic performance.