Bidirectional feature matching based on deep pairwise contrastive learning for multiparametric MRI image synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Multi-parametric MRI synthesis aids diagnosis when specific modalities are unavailable.
• Technical limitations often prevent acquiring all necessary MRI modalities for a patient.

Purpose of the Study:

• To propose a novel multi-parametric MRI synthesis model for generating target MRI modalities from two available ones.
• To enhance pathological MR image analysis by synthesizing missing contrasts.
• To improve diagnostic capabilities in scenarios with limited MRI acquisition.

Main Methods:

• A contrastive learning approach trains an encoder for target space feature extraction.
• A synthesis network generates target images from a common feature space.
• Bidirectional feature learning and a combined reconstruction and bidirectional triplet loss are employed.

Main Results:

• The model achieved average improvement rates of 3.9% (IXI dataset) and 3.6% (BraTS'18 dataset) over state-of-the-art methods.
• On the BraTS'18 dataset, the model recorded the highest Dice score of 0.793(0.04) for synthesized tumor regions.
• The model demonstrated flexibility in synthesizing head and neck CT images from MR acquisitions.

Conclusions:

• The proposed model efficiently generates diverse MR contrasts and preserves tumor areas in synthesized images.
• The model's flexibility extends to cross-modality synthesis (e.g., MR to CT).
• Future work includes validation for interventional MRI in neurosurgery and radiotherapy applications.