Medical Image Synthesis with Deep Convolutional Adversarial Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical imaging is crucial but limited by cost and radiation.
• Medical image synthesis offers a solution by generating images without new scans.
• Existing methods struggle with realism and accuracy in synthesizing medical images.

Purpose of the Study:

• To develop an accurate and robust method for medical image synthesis using generative adversarial networks.
• To generate realistic target images (e.g., CT from MRI, 7T MRI from 3T MRI) without additional scanning.
• To improve upon state-of-the-art methods in medical image synthesis tasks.

Main Methods:

• A fully convolutional network (FCN) was trained for image-to-image translation.
• Generative adversarial learning was employed to enhance the realism of synthesized images.
• An image-gradient-difference loss function and long-term residual units were incorporated to prevent blurriness and aid training.
• An Auto-Context Model was used to implement a context-aware deep convolutional adversarial network.

Main Results:

• The proposed method accurately and robustly synthesizes target medical images from source images.
• Evaluated on three datasets, the method successfully generated CT from MRI and 7T MRI from 3T MRI images.
• The method outperformed state-of-the-art comparison methods across all evaluated datasets and tasks.

Conclusions:

• The developed generative adversarial approach is effective for medical image synthesis.
• The method offers a promising solution for overcoming limitations in medical image acquisition.
• This technique advances the field of medical image synthesis, providing accurate and realistic results.