Abstracts

Rationale: Electrical source imaging (ESI) is a technique that estimates the location of sources of electrical current responsible for scalp potentials as measured by EEG electrodes positioned on the scalp. ESI is obtained by formulating a forward model that describes how electrical potential generated by the source propagates to the scalp, and solving an inverse problem that provides mapping of measured scalp potentials to estimated sources. Precise brain tissue segmentation and electrical conductivity values are necessary for an accurate forward solution. As a result, ESI is moving from generic and conformal atlases to patient-specific head models derived from individual MRIs. This work presents a novel approach for fully automatic, rapid and accurate head tissue segmentation that enables clinically efficient ESI. Methods: A shape-constrained deformable head model consisting of triangular meshes representing the scalp, skull, eyeballs, and air cavities was developed as an extension of previous work. Feature functions of the traditional shape-constrained deformable framework were replaced with boundary detectors based on convolutional neural networks (CNN). To complete the set of tissue types needed for ESI, voxel-wise image segmentation using labeled ground truth data from a publicly available repository (www.mindboggle.info) was applied to bootstrap and train a CNN-based classifier for brain tissues such as white matter (WM), grey matter (GM), and cerebrospinal fluid (CSF). Results: The adaptation of the shape-constrained head model to a new MRI scan is fully automatic. It takes less than a minute on a standard laptop (HP Zbook 15, 8GB RAM). An illustration of the surface representation of the head model is provided in Figure 1. An example of the model before and after adaptation to a new scan is shown in Figure 2. Cross-validation with manually expert-traced ground truth data was used to evaluate the accuracy of segmentation quantitatively. The average surface distance between the adapted model and the known ground truth was estimated for each tissue type. Volume overlap as defined by the Dice's coefficient with respect to the ground truth was used to evaluate the accuracy of the surface-based segmentation and the subsequent voxel-wise tissue classification. Results are illustrated in Figure 3. Conclusions: Fully automatic, rapid and highly accurate segmentation of head tissue types would enable the clinical use of ESI and ensure greater confidence in source localization. Results presented in this work suggest that the combination of a shape-constrained deformable segmentation followed by a machine learning-based voxel-wise tissue classification have the necessary accuracy to segment an individual MRI in a clinically acceptable time, and more importantly, support a real-world clinical workflow. Coupled with the high temporal resolution of EEG, ESI using this highly accurate segmentation has the potential to greatly enhance our understanding of underlying disease mechanisms and their related clinical manifestation. Funding: This study was funded by Philips Neuro.