Biologically Inspired Medical Multi-Modal Dataset Distillation via Contrast-Aware Alignment and Memory Compression

Area of Science:

• Artificial Intelligence
• Neuroscience
• Medical Imaging

Background:

• Multi-modal Magnetic Resonance Imaging (MRI) offers valuable diagnostic information but faces challenges in storage, privacy, and annotation costs.
• Biological vision systems efficiently integrate multi-sensory data and compress experiences for memory.
• Existing methods struggle with the scale and privacy concerns of large medical imaging datasets.

Purpose of the Study:

• To develop a bio-inspired framework for efficient and privacy-preserving compression of multi-modal MRI datasets.
• To mimic biological principles of perception and memory consolidation for data compression.
• To reduce storage requirements and annotation costs for large-scale medical imaging data.

Main Methods:

• Proposed a novel framework: Contrast-Guided Multi-modal Dataset Distillation (CGMDD).
• Employed hierarchical cross-modal contrastive learning for perceptual alignment across MRI modalities.
• Utilized dynamic dataset distillation mimicking memory consolidation for data compression via gradient-based optimization.

Main Results:

• CGMDD compressed datasets to 5% of their original size.
• The framework maintained competitive performance with only 30% of the original labels.
• Demonstrated effective integration of perception and learning through bio-inspired mechanisms.

Conclusions:

• Bio-inspired mechanisms are effective for creating efficient, robust, and privacy-preserving computer vision systems.
• CGMDD offers a promising solution for managing large-scale multi-modal MRI data.
• The framework highlights the potential of integrating neuroscience principles into AI for medical imaging.