MMSupcon: An image fusion-based multi-modal supervised contrastive method for brain tumor diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Accurate brain tumor diagnosis is crucial for effective treatment.
• Magnetic Resonance Imaging (MRI) is a key non-invasive diagnostic tool.
• Integrating multi-modal MRI data is essential for pattern recognition in tumor diagnosis, but limited by sample scarcity.

Purpose of the Study:

• To propose a novel training paradigm, multi-modal supervised contrastive learning (MMSupcon), to address the challenge of limited multi-modal imaging samples in brain tumor diagnosis.
• To enhance diagnostic accuracy by effectively fusing complementary MRI modalities and optimizing feature learning.

Main Methods:

• Developed a multi-modal medical image fusion component to create information-rich samples.
• Introduced a multi-modal supervised contrastive loss to guide feature learning, preserving cross-modal integrity and modality distinctiveness.
• Validated the MMSupcon method on a real-world brain tumor dataset from Beijing Tiantan Hospital and two public BraTS glioma classification datasets.

Main Results:

• MMSupcon achieved state-of-the-art performance on the real-world dataset.
• Demonstrated substantial performance improvements on two public BraTS glioma classification datasets.
• The proposed method effectively enhances diagnostic accuracy in multi-modal brain tumor imaging.

Conclusions:

• MMSupcon offers a significant advancement in multi-modal brain tumor diagnosis.
• The method successfully overcomes the limitations of scarce multi-modal imaging samples.
• The approach holds promise for improving clinical diagnostic accuracy and patient outcomes.