CD-GANs : Conditional dynamic training with GANs for enhanced post-gadolinium glioblastoma MRIs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Gliomas are aggressive brain tumors requiring accurate monitoring via MRI.
• Gadolinium contrast agents used in MRI pose risks, especially for patients with renal issues.
• There is a critical need for safer, effective alternatives for glioma imaging.

Purpose of the Study:

• To develop a method for synthesizing T1-weighted Contrast-Enhanced (T1CE) MRI images without gadolinium.
• To improve glioma visualization and characterization, reducing patient risks.
• To enhance diagnostic accuracy and treatment monitoring for brain tumors.

Main Methods:

• A Generative Adversarial Network (GAN) framework was developed for synthetic T1CE MRI generation.
• Dynamic training with condition-based discriminators and progressive intensity training were employed.
• The method was applied to Pix2pix, CPD-GAN, and DCD-GAN, creating PiPix2pix, PiCPD-GAN, and PiDCD-GAN.

Main Results:

• The proposed GAN models, particularly DCD-GAN and its progressive variant (PiDCD-GAN), demonstrated superior performance in T1CE image synthesis.
• The methodology showed effectiveness in multi-modality imaging on the Brats2020 dataset.
• Synthetic T1CE images accurately delineated tumor characteristics, comparable to traditional methods.

Conclusions:

• The developed GAN-based approach offers a safe and reliable alternative for T1CE MRI synthesis in glioma patients.
• This innovation can significantly reduce gadolinium-related risks while maintaining diagnostic quality.
• The study paves the way for improved glioma management through advanced, non-invasive imaging techniques.