A microstructural neural network biomarker for dystonia diagnosis identified by a DystoniaNet deep learning platform (Valeriani and Simonyan, PNAS, 2020).

The diagnostic accuracy of dystonia is unreliable, and the consensus between physicians is generally hard to achieve. The development of an objective machine learning-based objective diagnosis of dystonia would be imperative in significantly reducing the burden of the physician’s rate of mis- and under-diagnosis, improving the patient’s quality of life, and greatly decreasing the overall costs associated with delayed treatment of this disorder. Funded by the National Institute on Deafness and other Communication Disorders, National Institutes of Health (NIDCD/NIH R01DC011805), Amazon Web Services, and Mass General Brigham Innovation, our current studies combine large-scale multimodal neuroimaging (structural and functional MRI) with advanced deep-learning and shallow machine-learning algorithms to develop objective and accurate diagnostic markers of spasmodic dysphonia. We have developed the first objective, accurate, fast and cost-efficient deep-learning platform, DystoniaNet​​​​​​​, for dystonia diagnosis, which is currently being tested for its translation into clinical setting.