A UNIFIED CONDITIONAL DISENTANGLEMENT FRAMEWORK FOR MULTIMODAL BRAIN MR IMAGE TRANSLATION

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Multimodal Magnetic Resonance Imaging (MRI) offers crucial insights into tissue health and disease but acquiring multiple MRI sequences from a single subject is often challenging.
• Quantitative analysis of medical images is essential for accurate diagnosis and treatment planning, yet data limitations can hinder this process.

Purpose of the Study:

• To develop a unified framework capable of synthesizing any arbitrary MRI modality from a given input modality.
• To address the limitations of acquiring complete multimodal MRI datasets for comprehensive analysis.

Main Methods:

• A cycle-constrained conditional adversarial training approach was employed, featuring a modality-agnostic encoder for extracting invariant anatomical features.
• A conditioned decoder was utilized to generate the target MRI modality.

Main Results:

• The framework was validated on four MRI modalities (T1-weighted, T1 contrast-enhanced, T2-weighted, FLAIR) using the BraTS'18 database.
• The proposed method demonstrated superior synthesis quality compared to existing approaches.
• Experiments on a tumor segmentation task using synthesized data showed promising results.

Conclusions:

• The developed framework effectively synthesizes missing MRI modalities, overcoming data acquisition limitations.
• This approach has the potential to improve quantitative analysis and downstream tasks like tumor segmentation in medical imaging.