Joint self-supervised and supervised contrastive learning for multimodal MRI data: Towards predicting abnormal...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Multimodal magnetic resonance imaging (MRI) data integration with deep learning shows promise for disease diagnosis.
• Current methods struggle with fusing heterogeneous multimodal features effectively, leading to redundancy and loss of complementary information.
• Robust feature representation is crucial for leveraging the full potential of multimodal MRI.

Purpose of the Study:

• To develop a novel joint self-supervised and supervised contrastive learning method for multimodal MRI data.
• To learn robust latent feature representations by projecting heterogeneous features into a shared common space.
• To amalgamate complementary and analogous information across different MRI modalities and subjects for improved analysis.

Main Methods:

• A joint self-supervised and supervised contrastive learning framework was designed.
• The method projects heterogeneous multimodal MRI features into a shared latent space.
• Comparative analysis was performed against alternative deep multimodal learning approaches.

Main Results:

• The proposed method significantly outperformed several other deep multimodal learning techniques.
• Experiments were conducted on two independent datasets, demonstrating superior performance.
• The approach proved effective in predicting abnormal neurodevelopment.

Conclusions:

• The novel contrastive learning method offers a superior approach for multimodal MRI data fusion.
• This technique can enhance the accuracy of computer-aided diagnosis in clinical practice.
• The method effectively harnesses the power of multimodal data for improved neurodevelopmental disorder prediction.