Deep Learning to Construct Computational Head Models for Modelling Electroconvulsive Therapy

Abstract

Electroconvulsive therapy (ECT) is a neuromodulatory technique used widely for treatment of various psychiatric diseases such as Alzheimer’s and epilepsy. It includes the application of a large dose of electric current for short periods of time through attached electrodes on the scalp. However, there are some concerns related to electric field (EF) distribution in the brain and the potential for cognitive side effects. These concerns can be investigated thoroughly using computational modelling via the finite element (FE) method. To construct FE head models for simulating ECT stimulation, segmentation of magnetic resonance image (MRI) head scans is essential. This segmentation can be performed manually, which is laborious and time-consuming, or automatically. Various automatic segmentation approaches have been utilized to segment MRI head scans, including intensity-based techniques and deep learning techniques. In particular, convolutional neural networks (CNN), a subfield of deep learning, are considered as state-of-the-art owing to their outstanding performance. In this research, 2D U-Net and 3D U-Net CNN architectures were employed to rapidly segment seven MRI head scans into not only cerebrospinal fluid (CSF), grey matter (GM) and white matter (WM) as most studies performed but to include skin, skull tissues. This in turn will accelerate the process of construction of FE head models for simulating ECT. Two experiments were conducted using 2D U-Net, where in experiment 1, all the extracted slices from each head scan were used for training and testing, while experiment 2 involved only mid-slices. Additionally, two experiments were carried out using 3D U-Net with different patch sizes, 64 × 64 × 64 (experiment 3) and 128 × 128 × 128 (experiment 4). Inverse frequency class weighting was also used as an additional run for each experiment. The segmentation performance was evaluated in terms of intersection over union (IOU). In general, higher IOU scores were obtained for experiment 2 and experiment 4. The process of 3D segmentation (experiment 4) took only several seconds compared with previous approaches, which would take few hours. Predicted outputs from experiment 4 were used to construct FE models to simulate ECT. Various electrode montages were modelled on the scalp of each constructed head model, and EF magnitudes in various brain regions under ECT were compared against the corresponding manually-segmented head models. Larger EF magnitudes (10-25%) were exhibited in the automatically-segmented head models of four subjects, while weaker EFs (5-7%) were exhibited in the other three subjects compared with manual segmentation. These errors could be improved by increasing the training dataset to more than seven head scans, enhancing the segmentation accuracy for more accurate simulation of ECT.