SEMPAI: a Self-Enhancing Multi-Photon Artificial Intelligence for Prior-Informed Assessment of Muscle Function and...

Area of Science:

• Biomedical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) in biomedicine often requires large datasets and lacks interpretability.
• Fundamental laboratory research frequently deals with small, sparse data and existing biological knowledge.
• Current DL models can be black boxes, hindering expert involvement and knowledge discovery.

Purpose of the Study:

• To develop a DL framework for hypothesis-driven laboratory research using multiphoton microscopy (MPM).
• To enable DL models to integrate prior biological knowledge and provide interpretable feedback.
• To address data scarcity in fundamental research through meta-learning and multi-task learning.

Main Methods:

• Introduction of Self-Enhancing Multi-Photon Artificial Intelligence (SEMPAI).
• Utilizes meta-learning for simultaneous optimization of prior integration, data representation, and network architecture.
• Employs multi-task learning to enhance predictive performance on limited datasets.

Main Results:

• SEMPAI facilitates hypothesis testing with DL and offers interpretable insights into 3D image data.
• Applied to a decade-long MPM database of single muscle fibers, achieving the largest joint analysis of pathologies and function.
• Outperformed state-of-the-art biomarkers in six out of seven prediction tasks, particularly those with scarce data.

Conclusions:

• SEMPAI enables effective DL application in data-scarce, hypothesis-driven biomedical research.
• Integrated prior knowledge significantly improves DL model performance compared to models without priors or prior-only approaches.
• The developed method enhances knowledge discovery and interpretability in MPM-based laboratory studies.