Objective: A pilot study using an AI-based framework to classify routine brain scans as Parkinson’s disease (PD) or control.

Background: Prospective neuroprotective treatments for PD are highlighting the need for early diagnostic tests. MRI is not currently considered a robust imaging test for PD, but exploratory techniques have suggested that dedicated experimental sequences may be able to detect early pathological brain changes. We explored whether such changes might be detectable in routine MRI scans by employing deep learning (DL) methods. This subset of machine learning has recently shown great promise in diagnostic medical imaging, with its potential to detect patterns invisible to the human eye. Emerging explainability methods are allowing DL predictions to be better interpreted.

Method: T2 axial images were acquired from the Parkinson’s Progression Marker Initiative. Convolutional neural networks were developed to classify scans as PD or control. Models were developed for four PD cohorts stratified by time since diagnosis: 194 scans acquired more than four years post-diagnosis, 265 acquired two-to-four years post-diagnosis, 241 acquired one-to-two years post-diagnosis, and 282 acquired less than a year post-diagnosis. Each cohort was matched with controls based on age and sex. Models were assessed using 10-fold cross-validation. SHapley Additive exPlanations (SHAP) were used to visualize the contribution of individual pixels to the prediction.

Results: The classification performances for the developed models are reported in Table 1. The greater the time since diagnosis, the better the diagnostic performance of the models. The SHAP heatmaps for all cohorts demonstrated predominant contribution to the classification in the midbrain slices, as seen in Figure 1. [table1] [figure1]

Conclusion: The models trained on the later cases of PD exhibited good diagnostic performance. The decreasing performance for earlier stages of PD suggests that progressive changes have been detected. The use of explainable AI highlighted regions of interest consistent with the known neuropathology of PD, providing a focus for future work. We aim to validate the results of this pilot study in a large dataset comprised of routine brain imaging from National Health Service patients in the South West of England.

« Back to 2023 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/explainable-deep-learning-based-detection-of-parkinsons-changes-in-mri-brain-scans/