Editorial: Machine learning methods for human brain imaging

Fatos Tünay Yarman Vural, Sharlene D. Newman, Tolga Çukur, Itir Önal Ertugrul

Abstract: The research topic, Machine Learning Methods for Human Brain Imaging, is a small sampling of 11 research articles that demonstrate the use of ML in multiple contexts and with multiple imaging modalities. The collection includes two manuscripts (Alchihabi et al., 2021;Fang et al., 2022) that take different approaches to understanding cognitive networks -one using multi-variate pattern dependencies between brain regions and another examining network dynamics during the execution of a task. There are also three studies designed to use AI to diagnose psychological disorders -one using MRI to diagnosis defiant disorders in children (Menon and Krishnamurthy, 2021), one using EEG to classify brain states in schizophrenia patients and healthy controls (Plechawska-Wójcik et al., 2021) and another classifying patients with obsessive-compulsive disorder and controls (Luo et al., 2021). A third group of studies use ML to address analytic issues including one developing an open access tool for whole brain segmentation (Manjón et al., 2022) and volumetric analysis of large datasets, one using fuzzy neural networks to improve 2D to 3D image transformations (Tavoosi et al), and registration of multimodal 2D coronal section images of gene expressions in the mouse brain (Krepl et al., 2021).One goal of AI is to create systems that function like the human brain (Hopgood, 2005). Current systems fall short and two of the manuscripts in this collection attempt to address this issue (Matsui et al., 2022;Zhang et al., 2021). Deep learning systems, for example do a great job of mimicking human vision, to a point; their mapping from stimulus input to perceptual output are different with respect to adversarial images. Zhang et al attempts to characterize the differences in how AI systems and human brains process these adversarial images by comparing artificial neural networks and human brain activation, using what is learned to improve AI performance.The use of ML in human brain imaging is only expected to increase. The power of deep learning methods makes them attractive for analyzing the growing number of large, publicly available datasets. However, it is important to slow down to evaluate their efficacy as well as to evaluate their weaknesses. One such weakness is addressed in the manuscript by Varotto et al.2021how to handle imbalanced datasets. Most large datasets do not have even distributions of minority populations (e.g., racial, socioeconomic, patient, etc.). This is only one such shortcoming that demonstrates the need for careful evaluation.