Applications of deep learning in magnetic resonance imaging brain disease diagnosis

In recent years, with the development of deep learning, the impressive achievements of deep learning in several fields, such as computer vision and natural language processing, have shown its great potential for development , as well as gradually having an impact on the field of brain disease diagnosis. The development of deep learning in aid of diagnosis of brain diseases has become increasingly effective. In this article, we overview several deep learning-based approaches that have been frequently used in the past three years to aid in the diagnosis of typical brain disorders, including convolutional neural networks that preserve the spatial structure of input features, recurrent neural networks that take into account correlations between input feature sequences, and graph convolutional neural networks that process features well in non-Euclidean spaces. We categorize these studies by disease to reflect succinctly the recent developments in deep learning in aid of diagnosis of brain diseases. The structure of this review is as follows: we first overview and summarize the development of some mainstream general frameworks during the history of deep learning, then we introduce the implementation of these methods in the aid of diagnosis of typical brain diseases, and finally we conclude with a summary and outlook on the implementation of deep learning in the aid of diagnosis of brain diseases. This review covers three typical brain disorders, each containing about 15-20 articles, and these brain disorders are Autism Spectrum Disorder (ASD), Schizophrenia (SZ), and Alzheimer's Disease (AD), autism is a neurodevelopmental disorder that occurs in early childhood, schizophrenia is a psychiatric disorder that occurs in young adulthood, and the last one is Alzheimer's disease that occurs in old age. In the application of deep learning methods to the three brain disorders, we classify them according to the characteristics of the different inputs. Most of the literature in which MR images are used directly as input uses convolutional neural networks as the backbone network for designing the further feature extraction methods. When dealing with data containing sequence information with many time points, recurrent neural networks are used to extract key information among the sequences. Due to the huge amount of data, in addition to processing the image directly as an input, a number of papers have also taken the approach of extracting manual features and furthermore constructing graph structures for manual features, and analyzing them using graph neural network-based methods, which also obtained good results. At the same time, it can be seen that the graph neural network-based analysis method has become a great trends.