Generative Adversarial Networks Can Create High Quality Artificial Prostate Cancer Magnetic Resonance Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning in medicine requires specific data, often limited by disease heterogeneity.
• Generative Adversarial Networks (GANs) offer a potential solution for creating synthetic medical data.
• Prostate cancer's heterogeneity poses challenges for current GAN models.

Purpose of the Study:

• To train GAN models using prostate MRI images to generate synthetic data.
• To validate the quality and clinical utility of the generated synthetic images.
• To assess the potential of synthetic MRI data for machine learning applications in oncology.

Main Methods:

• T2-weighted prostate MRI images from the BLaStM trial were used to train Single Natural Image GANs (SinGANs).
• A deep learning semantic segmentation pipeline was employed for prostate boundary segmentation.
• Synthetic images underwent quality control assessment by scientists with varying experience levels and a blinded radiologist, and were used for anomaly detection.

Main Results:

• Experienced scientists achieved 67% accuracy in distinguishing real from synthetic images; less experienced groups performed at 58% and 50%.
• Nearly half (47%) of synthetic images were misclassified as real.
• A radiologist could not significantly differentiate between real and synthetic images in a blinded assessment.
• Anomaly detection success rates were comparable for real and synthetic images.

Conclusions:

• GANs can generate high-quality synthetic prostate MRI images that are difficult to distinguish from real images.
• Synthetic data generated by GANs show potential for use in clinical applications involving supervised machine learning.
• This AI-driven approach may significantly contribute to advancing medical imaging analysis and disease study.