Breast Ultrasound Image Synthesis using Deep Convolutional Generative Adversarial Networks.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Image Synthesis
• Radiology and Ultrasound Technology

Background:

• Deep convolutional generative adversarial networks (DCGANs) are advanced AI tools for generating synthetic images.
• The clinical utility of synthesized medical images, particularly in breast ultrasound, requires thorough evaluation.
• Assessing the quality and realism of AI-generated images is crucial for their adoption in healthcare.

Purpose of the Study:

• To evaluate the clinical utility of synthesized breast ultrasound images generated by DCGANs.
• To assess the quality, definition of anatomical structures, and visualization of masses in synthesized images.
• To determine the indistinguishability of synthesized images from original breast ultrasound images.

Main Methods:

• Retrospective collection of 528 benign and 529 malignant breast mass ultrasound images.
• Generation of synthesized images using a DCGAN at varying epochs (50, 100, 200, 500, 1000).
• Evaluation of synthesized (n=20) and original (n=40) images by two radiologists using a five-point scale for quality, anatomical definition, mass visualization, and realism.

Main Results:

• Images synthesized with 500 epochs were rated highest in quality compared to other synthesized image sets.
• A significant percentage of synthesized images (up to 37.5% for 500 epochs) were indistinguishable from original images.
• Interobserver agreement among radiologists was very good, indicating reliable image assessment (|r|=0.708-0.825, p<0.001).

Conclusions:

• DCGANs can generate high-quality, realistic synthesized breast ultrasound images.
• The synthesized images demonstrate significant clinical value and realism, comparable to original images.
• This technology holds promise for enhancing medical imaging datasets and training applications.