A physics-informed geometric deep learning method for electrophysiological source reconstruction

Thomas M. Menino Convention & Exhibition Center 

Electrophysiological brain signals are typically acquired through indirect and noisy measurements, providing transformed representations of the underlying neural activity. Source reconstruction --- the inverse problem of resolving underlying neural signals from these measurements --- is essential for accurate brain function mapping. However, this task remains extremely challenging due to its mathematical ill-posedness and resulting sensitivity to noise. Deep learning methods have shown promise across a range of inverse problems, but they often disregard the underlying physical principles governing the data generation process, leading to inefficient learning. In this talk, I will introduce a novel physics-informed geometric deep learning framework that embeds potentially ill-posed physics constraints into the model via a custom layer, resulting in more efficient learning and improved reconstruction performance. This custom layer enables the neural network to adapt to subject-specific variations in the physics of signal generation and, as a byproduct, to seamlessly handle missing data within the sensor measurements. We demonstrate the proposed method on magnetoencephalography source reconstruction from a cohort of adolescents, which is characterized by particularly low signal-to-noise ratios.

brain function

Deep learning

neural network