Personalized Deep Learning based Source Imaging Framework Improves the Imaging of Epileptic Sources from MEG Interictal Spikes

Abstract

Electromagnetic source imaging (ESI) has been widely used to image brain activities for research and clinical applications from MEG and EEG. It is a challenging task due to the ill-posedness of the problem and the complexity of modeling the underlying brain dynamics. Deep learning has gained attention in the ESI field for its ability to model complex distributions and has successfully demonstrated improved imaging performance for ESI. In this work, we investigated the capability of imaging epileptic sources from MEG interictal spikes using deep learning-based source imaging framework (DeepSIF). A generic DeepSIF model was first trained with a generic head model using a template MRI. A fine-tuning procedure was proposed to introduce personalized head model information into the neural network for a personalized DeepSIF model. Two models were evaluated and compared in extensive computer simulations. The MEG-DeepSIF approach was further rigorously validated for imaging epileptogenic regions from interictal spike recordings in focal epilepsy patients. We demonstrated that DeepSIF can be successfully applied to MEG recordings and the additional fine-tuning step for personalized DeepSIF can alleviate the impact of head model variations and further improve the performance significantly. In a cohort of 29 drug-resistant focal epilepsy patients, the personalized DeepSIF model provided a sublobar concordance of 93%, sublobar sensitivity of 77% and specificity of 99%, respectively. When compared to the seizure-onset-zone defined by intracranial recordings, the localization error is 15.78 +- 5.54 mm; and when compared with resection volume in seizure free patients, the spatial dispersion is 8.19 +- 8.14 mm. DeepSIF enables an accurate and robust imaging of spatiotemporal brain dynamics from MEG recordings, suggesting its unique value to neuroscience research and clinical applications.

Spatiotemporal Rhythmic Seizure Sources Can be Imaged by means of Biophysically Constrained Deep Neural Networks

“…They have shown excellent performance in computer simulations, demonstrating the power of DL-based ESI methods. DeepSIF as a DL-based ESI method, has proven to be effective for imaging transit activities such as interictal spikes or evoked potentials in a large group of subjects [14,15]. It is a modular framework consisting of a forward source model, using neural mass models, to generate realistic synthetic training data, and an inverse neural network model to perform the ESI task based on the information in the training data.…”

“…The Otsu's thresholding technique [55] was used to identify the boundary of the imaged source distribution. Modified spatial sensitivity and specificity [14,56] were used to evaluate the extent estimation accuracy. The localization error (LE) is defined as the average of the distance from the estimated source to the ground truth and the distance from the ground truth to the estimated source [57].…”

“…Computational models can serve as a powerful alternative to introduce the source dynamics into training data, if the source models are realistic enough for source imaging tasks. DL-based source imaging framework (DeepSIF) is recently proposed as a general ESI framework unitizing dynamic brain network models as the source model in the forward problem, and the deep neural networks (DNN) to solve the inverse problem [14,15]. Different source models can be selected based on source signal properties and the modalities of the measurements, to generate various types of source-sensor signal pairs as the training data.…”