Deep learning-based multi-modal computing with feature disentanglement for MRI image synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Different Magnetic Resonance Imaging (MRI) modalities provide unique pathological information.
• Acquiring full-sequence MRI scans is challenging due to time and cost constraints.
• Accurate prediction of target MRI sequences is crucial for comprehensive clinical diagnosis.

Purpose of the Study:

• To develop a highly accurate algorithm for predicting target MRI sequences.
• To provide additional information for clinical diagnosis by synthesizing missing MRI modalities.
• To overcome limitations of time consumption and high cost associated with obtaining complete MRI scans.

Main Methods:

• A deep learning-based multi-modal computing model for MRI synthesis is proposed.
• Feature disentanglement strategy separates modality-invariant and modality-specific information.
• Adaptive Instance Normalization (AdaIN) and a Local Adaptive Fusion (LAF) module are employed for effective feature fusion and pseudo-target generation.

Main Results:

• The method was validated on the BRATS2015 dataset, demonstrating superior performance over benchmark and state-of-the-art methods.
• Quantitative measures showed significant improvements, with Peak Signal-to-Noise Ratio (PSNR) increasing from 23.68 to 24.8 compared to pix2pixGANs.
• Ablation studies confirmed the efficacy of key components within the proposed model.

Conclusions:

• The developed method is effective for predicting target MRI sequences.
• The algorithm offers valuable support for clinical diagnosis and treatment planning.
• This approach enhances the utility of MRI in medical practice by enabling synthesis of crucial sequences.