Model-Based Explainable Deep Learning for Light-Field Microscopy Imaging.

Area of Science:

• Neuroscience
• Computational Imaging
• Biophysics

Background:

• Understanding neural network dynamics requires observing large neuronal populations.
• Light-field microscopy (LFM) offers high-speed, 3D imaging for this purpose.
• Existing LFM computational methods need improved interpretability and transparency.

Purpose of the Study:

• To develop a model-based explainable deep learning approach for LFM.
• To integrate domain knowledge from physics and optics into neural network models.
• To enhance the performance, interpretability, and transparency of LFM data analysis.

Main Methods:

• Proposed a novel deep learning architecture integrating wave-optics, sparse representation, and non-linear optimization.
• Employed a hybrid training strategy combining layer-wise training and knowledge distillation.
• Utilized a model-based explainable AI framework for LFM data processing.

Main Results:

• Achieved fast and robust 3D localization of neuron sources in scattering mammalian brain tissue.
• Demonstrated accurate identification of neural activity using LFM data.
• The integrated model-based and learning-based approach yielded superior performance.

Conclusions:

• The proposed explainable deep learning method significantly advances LFM data analysis.
• This approach offers a powerful tool for neuroscience research by improving neural activity observation.
• The hybrid method provides a transparent and interpretable framework for complex biological imaging.