For clinical management and treatment planning to be successful, Parkinson’s disease (PD) and its early stages must be identified promptly and accurately. This study introduces a lightweight two-block convolutional neural network (CNN) that uses 2D brain MRI slices to perform early-stage stratification (Stage 1 vs. Stage 2) and binary classification (PD vs. healthy). The model is composed of five convolutional layers with ReLU activation and max pooling, a fully connected layer with 1,024 units, and a SoftMax output layer. It is trained on 831 labelled MRI images with a 50 x 50-pixel size. The Adam optimizer (learning rate = 1 × 10⁻3), a batch size of 5, and 100 epochs were used for training. On the training set, the model’s accuracy was 100%, and on a held-out test, it was 96% accuracy on a held-out test set (n = 25), with F₁-scores of 0.94 (precision) and 0.89 (recall) for PD classification and an AUC of 1.00. The test cohort’s disease stages were perfectly classified when a threshold of 0.94 was applied to the PD class probability for stage prediction. Now the system is now an interactive Streamlet web application that enables real-time inference from single MRI slice uploads, producing diagnosis, stage, and confidence scores to improve accessibility, with no requirement for artificial data or intensive preprocessing. Our method provides a good balance between accuracy and computational efficiency when compared to current MRI-based and multimodal approaches.