Artificial Intelligence based detection of Parkinson’s disease in Magnetic Resonance Imaging brain scans

Aims: Candidate neuroprotective treatments for Parkinson’s disease (PD) are highlighting the need for early diagnostic tests. Exploratory imaging techniques have suggested that early pathological brain changes may be detectable using dedicated experimental MRI sequences. We explored whether deep learning might be employed to detect such brain changes on routine MRI scans. Deep learning has shown promise in diagnostic medical imaging, and offers the potential of automated diagnosis by detecting patterns that might be invisible to the human eye. These methods have sometimes been criticised for being “black boxes”, but emerging explainability methods are allowing better interpretation. Methods: We trained a convolutional neural network to classify 138 PD and 60 control brain MRI images acquired from the Parkinson’s Progression Marker Initiative database. Models were assessed using 5-fold cross-validation. We used Deep SHapley Additive exPlanations (DeepSHAP) to calculate and visualise the contribution of individual pixels to the model’s prediction. Results: A model trained using a combined dataset of axial T2 and proton density MRI scans classified images with 79% accuracy and a Receiver Operating Characteristic area under the curve (AUC) of 0.86. Another model trained on just T2 scans classified images with 81% accuracy and AUC of 0.83. A further model trained on just proton density scans classified images with 84% accuracy and AUC of 0.89. The heatmaps generated using DeepSHAP demonstrated predominant contribution to the prediction in the midbrain slices. Conclusions: Our models exhibited good diagnostic performance. The explainability method highlighted regions of interest consistent with the known neuropathology of PD, providing a focus for future work. We will validate these models in a large dataset of routinely collected MRI scans from South West England, many of which precede the onset of motor symptoms.