Anatomic and Molecular MR Image Synthesis Using Confidence Guided CNNs

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neuro-oncology

Background:

• Limited data quantity and quality hinder AI applications in neuro-oncology.
• Previous work developed the Synthesis of Anatomic and Molecular MR image networks (SAMR) for post-treatment malignant gliomas.
• Need for improved methods to synthesize high-quality, multi-modal MR images for better diagnostics.

Purpose of the Study:

• To develop a confidence-guided SAMR (CG-SAMR) model for synthesizing multi-modal MR images from lesion contour information.
• To enhance data synthesis using a confidence measure for intermediate results.
• To adapt the model for training with unpaired data.

Main Methods:

• Implemented a confidence-guided module within the SAMR architecture.
• Synthesized multi-modal MR images including T1-weighted, Gd-T1, T2-weighted, FLAIR, and amide proton transfer-weighted (APTw) sequences.
• Extended the model to accommodate training with unpaired clinical data.

Main Results:

• The CG-SAMR model demonstrated superior performance compared to state-of-the-art synthesis methods.
• Experiments on real clinical data validated the effectiveness of the proposed approach.
• The method successfully synthesized multi-modal MR images, including molecular APTw sequences.

Conclusions:

• The confidence-guided SAMR approach significantly improves multi-modal MR image synthesis for neuro-oncology.
• This method addresses data limitations and enhances diagnostic capabilities for malignant gliomas.
• The developed model offers a promising tool for advancing data-driven approaches in clinical neuro-oncology.