A least square generative network based on invariant contrastive feature pair learning for multimodal MR image...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurosurgery

Background:

• Acquiring multiple magnetic resonance (MR) sequences during neurosurgery is vital for surgical planning and tumor resection.
• Timing constraints and limited MR sequence availability can hinder intraoperative decision-making.

Purpose of the Study:

• To develop an automated method for synthesizing additional MR contrasts from existing heterogeneous MR sequences.
• To alleviate timing constraints and improve intraoperative guidance in MR-guided neurosurgery.

Main Methods:

• Proposed a multimodal MR synthesis approach using a least squares Generative Adversarial Network (LSGAN).
• Incorporated an unsupervised contrastive learning strategy with a contrastive encoder to extract invariant representations.
• Integrated a novel perception loss term alongside reconstruction loss for generator training.

Main Results:

• The model achieved the highest Dice score ([Formula: see text]) on the BraTS'18 brain tumor dataset.
• Demonstrated superior performance compared to other multimodal MR synthesis methods, with low variability information ([Formula: see text]).
• Achieved a probability rand index score of [Formula: see text] and a global consistency error of [Formula: see text].

Conclusions:

• The developed model reliably synthesizes MR contrasts, enhancing tumor visualization in brain tumor datasets.
• Future work includes clinical evaluation for residual tumor segmentation in MR-guided neurosurgeries with limited MR contrasts.