Deep learning has achieved remarkable accuracy in diagnosing Parkinson’s disease (PD) using Single-Photon Emission Computed Tomography (SPECT) imaging. however, its lack of interpretability continues to hinder clinical trust and adoption. This thesis propose Concept Gradient Vectors (CGVs), a novel, post-hoc, concept-based explanation method that leverages the continuous biomarker, Striatal Binding Ratio (SBR), to elucidate the model’s decision-making process. SBR quantifies dopamine transporter availability in the striatum and gradually declines with disease progression, making it a clinically meaningful and inherently continuous indicator of neurodegeneration. Unlike conventional concept-based methods that rely on discrete categories or linear assumptions, CGVs employ Support Vector Regression (SVR) with a radial basis function (RBF) kernel to model nonlinear relationships between latent features and SBR values. By computing the local gradient of the SVR function, CGVs reveal how subtle variations in biologically grounded concepts affect model predictions. Our method offers both local, sample-specific insights and global, population-level trends through gradient analysis. Experiments on the Parkinson's Progression Markers Initiative (PPMI) dataset demonstrate that CGVs yield more faithful and fine-grained explanations than linear alternatives, while aligning with known patterns of dopaminergic degeneration. Beyond enhancing interpretability, CGVs also help reveal differences in model confidence behavior and latent representation structure―contributing to more transparent and clinically grounded AI systems.