Deep Learning in Neuroradiology: A Systematic Review of Current Algorithms and Approaches for the New Wave of Imaging Technology

Purpose To systematically review and synthesize the current literature and to develop a compendium of technical characteristics of existing deep learning applications in neuroradiology. Materials and Methods A Preferred Reporting Items for Systematic Reviews and Meta-Analyses systematic review was conducted through September 1, 2019, using PubMed, Cochrane, and Web of Science databases. A total of 155 articles discussing deep learning applications in neuroimaging were identified, divided by imaging modality, and characterized by imaging task, data source, algorithm type, and outcome metrics. Results A total of 155 studies were identified and divided into: MRI (n = 115), functional MRI (n = 19), CT (n = 9), PET (n = 18), and US (n = 1). Seven were multimodal. MRI applications were described in 74%, and 76 (49%) were tasked with image segmentation. Of the 155 articles identified in this study, 65 (42%) were tested on institutional data; only 16 were validated against publicly available data. In addition, 53 studies (34%) used a combined dataset of less than 100, and 124 (80%) used a combined dataset of less than 1000. Conclusion Although deep learning has demonstrated potential for each of these modalities, this review highlights several needs in the field of deep learning research including use of internal datasets without external validation, unavailability of implementation methods, inconsistent assessment metrics, and lack of clinical validation. However, the rapid growth of deep learning in neuroradiology holds promise and, as strides are made to improve standardization, generalizability, and reproducibility, it may soon play a role in clinical diagnosis and treatment of neurologic disorders. Supplemental material is available for this article. Keywords: Brain/Brain Stem, Convolutional Neural Network (CNN), Head/Neck, Supervised learning © RSNA, 2020