Learning-Based Head Model Similarity Estimation for MRI-Free Brain Imaging

Numerous imaging techniques can be used to investigate the brain and its behavior, each with its strengths and its weaknesses. Among these we highlight Magnetic Resonance Imaging (MRI) and Electroencephalography (EEG). MRI allows us to observe the physical architecture of the brain, it has great spatial resolution, but its low temporal resolution often limits the functional information it can provide. MRI machines are also expensive and quite cumbersome, to the point that most research centers do not possess their own. These machines are usually bought by large clinics to serve a variety of patients with different conditions. This can lead to long waiting lists and delays, not to mention heavy costs due to the inherent price charged by these institutions for the usage of the equipment. EEG is instead capable of characterizing the electrical behavior of the brain with a high temporal resolution, and at a comparatively low cost. Its uses are numerous, from diagnostics to research, and even helping people affected by lockdown syndrome. However, a crucial shortcoming of EEG is its poor spatial resolution, which can however be improved by exploiting the highly defined MRI scan of the patient’s head. In this thesis we have attempted to alleviate the cumbersome requirement of acquiring patient-specific head models to fully exploit EEG as an electrophysiologic imaging modality. We investigated the capabilities of a learning-based algorithm to provide an alternate, but electrophysiologically relevant, head model for EEG modeling. In other words, the algorithm provides the user with a surrogate MRI scan that closely matches the electrical characteristics of a subject’s head, given only readily accessible information regarding the patient (physical features, age, . . . ). To this end we investigated the possibility of using a deep learning architecture capable of estimating the electrical similarity between patients based only on this type of information. One of the deep learning architectures we have explored for this task is a Siamese Neural Network (SNN) based on PointNet. A SNN works by taking two inputs, extracting their features and comparing them to produce a similarity score. It can be broken down into two main components: a network that handles the two inputs and reduces them to simple values on a vector, and a fully connected layer at the end to compare the features produced by the first network. In our case, the inputs would be external physical features of the subjects. It is possible to process 3D models in multiple ways inside a neural network, like using voxelization or directly mesh data, and, since this is still an active field of research, there is no widespread consensus on the best architectures. We have chosen to encode the models as points clouds since that is how it is encoded in the raw data from our scanner. We therefore based the feature extraction portion of the net on PointNet, a neural network developed by Stanford University for classification and segmentation of 3d objects represented as point clouds. We based our choice on its good performance in benchmarks and the relatively wide adoption the network received, even on similar tasks as ours. The selected scheme shows promising results for the relatively small dataset that were explored, further studies on large MRI datasets will be required to confirm that the scheme maintains these performances for more realistic scenarios and to verify if it suffers from biases.