Deep learning‐based scheme to diagnose Parkinson's disease

Abstract

AbstractParkinson's disease (PD) is a neurological disorder of the central nervous system that causes difficulty in movement, often including tremors and rigidity. Early detection of PD can prevent symptoms up to a certain age and increase life expectancy. For this purpose, we have used brain images from magnetic resonance imaging (MRI) technique. A deeper level of feature detection in MRI can identify biomarkers that can be used to know how the disease spreads, leading to a cure in the future. With these motives, we have presented two novel approaches using deep learning (DL) techniques. 2D and 3D convolution neural networks (CNN) are used, which are trained on MRI scans in the axial plane. The dataset was constructed using images from Parkinson's progression markers initiative (PPMI). The four pre‐processing techniques used in this article are bias field correction, histogram matching, Z‐score normalization, and image resizing. Pre‐processing techniques were essential inaccurate training models. Every class prediction done by the model would have taken multiple features into account across multiple layers of the brain and not relied on a single or few important features, making DL a powerful concept. A total of 318 MRI scans were used to train and test a 2D CNN and a 3D CNN model. We have compared the models' results using different evaluation parameters such as accuracy, loss, confusion matrix, receiver operating characteristic (ROC) curve, and precision‐recall (PR) curve. The 3D model learned key features from the data and was able to classify the test data with 88.9% accuracy with 0.86 area under curve (AUC). In contrast, the 2D model achieved a mediocre accuracy of 72.22% with 0.50 AUC. This shows that the 3D model is more accurate and reliable than the 2D model.