Early Detection of Parkinson’s Disease in T2-Weighted MRI Scans Using Grad-CAM

Parkinson’s disease (PD) is one of the most challenging diseases that severely affects patients’ quality of life. While the diagnosis is relatively straightforward once motor symptoms appear, at this stage, more than 60% of dopaminergic neurons have already been lost. Thus, early detection is essential to delay the progression and enable timely intervention. The treatment protocols are not able to cure the disease; however, they can alleviate symptoms and improve patient outcomes. This study proposes a deep learning network for the early detection of PD using T2-weighted MRI scans from the Progression Markers Initiative (PPMI) and the Alzheimer’s Disease Neuroimaging Initiative (ADNI) datasets. Moreover, it applies the Gradient-weighted Class Activation Mapping (Grad-CAM) explainability tool to analyze which regions participate the most in the model’s prediction. Magnetic Resonance Imaging (MRI) is an imaging technique that provides structural information about the brain. T2-weighted MRI is one of the MRI sequences that is particularly useful for the diagnosis of neurological disorders like Parkinson's disease. Since it highlights the brain tissue's contrasts relevant to these diseases. In this study, Image processing techniques are also applied to T2-weighted MRI scans to enable the extraction and enhancement of meaningful features from these scans. By combining imaging modalities with advanced processing methods, it becomes possible to detect subtle patterns associated with Parkinson’s disease. After attempting many CNN architectures, such as EfficientNet-B0, DenseNet, a ResNet-34 convolutional neural network was developed and optimized, and it achieved testing accuracy that exceeded 90%, which emphasizes its high ability to identify PD cases. This performance was further examined by applying Grad-CAM. The Grad-CAM results emphasize the ability of the trained model to not only determine Parkinson’s cases effectively, but also to learn patterns related to early biomarkers of PD, such as the occipital lobe, cerebellum, and frontal cortex. The high accuracy achieved by the model, together with the Grad-CAM results, indicates that this project has successfully achieved its objective of detecting Parkinson’s disease in its early stages, while also providing an explainable model for clinicians. The results ensure the importance of combining CNNs with explainability techniques for more reliable, non-invasive, and automated PD diagnosis. The future work should evaluate this model with larger datasets and attempt it with different imaging modalities to enhance generalization.