This study presents an AI-based system for early detection of Parkinson's disease using deep learning models Inception V3 and Xception with Attention Mechanism. The system analyzes hand-drawn spiral images, which serve as biomarkers for Parkinson's symptoms like tremors and micrographia. The proposed model extracts critical features from these images using pre-trained convolutional neural networks (CNNs) enhanced with attention layers, ensuring effective classification. The dataset includes spiral drawings from both healthy individuals and Parkinson's patients, allowing the model to learn distinguishing features. The Inception V3 model achieved 100% accuracy, while the Xception model attained 88% accuracy in Parkinson's detection. To evaluate the model's performance, graphs of accuracy against epochs and loss against epochs were plotted to track learning trends. A confusion matrix was generated to analyze misclassifications, and a classification report provided insights into precision, recall, and F1-score. A comparative bar chart was also used to highlight the performance difference between Inception V3 and Xception models. This AI-driven approach provides a non-invasive, cost-effective, and automated diagnostic tool, improving early diagnosis and assisting healthcare professionals in timely intervention.