Explainable Parkinson's Disease Detection from MRI Scans Using Pretrained ResNet-18

Parkinson's Disease (PD) is a degenerative neurological disease that affects millions around the world, and its early and accurate diagnosis from MRI scans can prove to be a boon in therapy planning and patient response. This work constitutes an explainable deep learning framework that is trained using a pretrained ResNet-18 model for the binary classification of PD versus healthy controls. The first convolutional layer is specially altered for adapting gray-scaled inputs, and the final, fully connected layer is replaced with a binary output. For more clinically transparent results, two model-agnostic interpretation techniques, SHAP for global as well as local pixel-level attributions, and LIME for superpixel-based explanations, are employed with it. Experimental results on benchmarks clearly show competitive results, achieving more than 90 % overall accuracy on the validation split. The visual explanations yield very similar brain regions concerning both dopaminergic loss and relevant PD pathology to those produced by SHAP and LIME. The performance of classification is also assured through confusion matrix analysis as well as standardized metrics-sensitivity and specificity. Our diagnostic performance at high levels is synergistic with clear, high-quality visual evidence for a trustworthy AI driving clinician-acceptable decision support in the detection of Parkinson's Disease.