Deep learning classification of MRI differentiates brain changes in genetic and idiopathic Parkinson's disease

BackgroundThe prospect of neuroprotective treatments for Parkinsons disease highlights the need for early diagnostic tests. Specialised MRI sequences suggest changes related to Parkinsons disease may be detectable. ObjectivesWe used the Parkinsons Progression Markers Initiative dataset to investigate whether deep learning can detect early brain MRI changes of idiopathic and GBA/LRRK2 prodromal Parkinsons disease. MethodsPairs of matched cohorts were used to train convolutional neural networks to classify T2 axial images. Explainability methods were used to visualise drivers of model predictions. ResultsModels built to distinguish between idiopathic Parkinsons disease scans (n=504) and matched controls exhibited good classification performance for scans taken more than four years after diagnosis, with a Receiver Operating Characteristic area-under-the-curve of 0.89 (n=98). Model performance deteriorated as time since diagnosis reduced. Models built to distinguish non-manifesting carriers of LRRK2 (area-under-the-curve 0.92, 90% accuracy, n=115) and GBA (area-under-the-curve 0.94, 92% accuracy, n=109) from controls exhibited excellent classification performance. All models demonstrated foci of interest in cerebrospinal fluid spaces surrounding the brainstem. Models using GBA scans also identified foci of interest in occipital lobes. ConclusionsDeep learning models appear able to reproducibly detect changes in the brains of those with established but not early Parkinsons disease. Conversely changes in at risk genetic cohorts are detectable at all stages, including in those who have not developed Parkinsons disease. This implies a distinct pathological process ongoing within the brains of carriers of Parkinsons disease genetic risk factors compared to those with sporadic Parkinsons disease.