Xeno-learning: knowledge transfer across species in deep learning-based spectral image analysis.

Area of Science:

• Biomedical engineering
• Medical imaging
• Machine learning

Background:

• Optical imaging techniques like hyperspectral imaging (HSI) show promise for surgical applications.
• Machine learning (ML) algorithms require large datasets, which are scarce in clinical settings.
• Preclinical animal data is abundant but difficult to apply directly to human patients due to ethical and biological differences.

Purpose of the Study:

• To introduce 'xeno-learning,' a novel cross-species knowledge-transfer framework.
• To demonstrate the feasibility of transferring ML models trained on animal data to human surgical imaging.
• To address the challenge of limited clinical data for ML-based surgical imaging.

Main Methods:

• Collected 14,013 hyperspectral images from human, porcine, and rat models.
• Developed a 'physiology-based data augmentation' method for cross-species knowledge transfer.
• Validated the transferability of learned spectral changes across species.

Main Results:

• Spectral signatures differ significantly across species, but relative pathological changes are comparable.
• Xeno-learning successfully transferred knowledge from preclinical models to human data.
• Physiology-based data augmentation enabled effective secondary use of animal data for human applications.

Conclusions:

• Xeno-learning offers a viable solution to the clinical data scarcity problem in ML-based surgical imaging.
• Cross-species knowledge transfer is achievable by focusing on relative physiological changes.
• This approach has the potential to significantly advance the development of AI-powered surgical tools.