| Parkinson's disease is a prevalent neurodegenerative disorder characterized by insidious onset, complex symptoms, and the accuracy of clinical diagnosis highly depends on the doctor's experience, lacking objective and quantitative diagnostic techniques. With the advancement of artificial intelligence, the integration of deep learning techniques holds promise as an accurate and efficient approach for automated PD diagnosis. This paper proposes a deep learning-based framework for assisting in the diagnosis of Parkinson's disease. Firstly, an image segmentation model is employed to segment the midbrain region of axial slices of the brain. Secondly, the MobilenetV2 network was improved by incorporating inception architecture, CA attention mechanism, and TanhExp activation function. The improved MobilenetV2 model was trained and tested using midbrain images, achieving a diagnostic accuracy of 97.5%, sensitivity of 97.53%, and recall of 97.48% in distinguishing PD cases from normal controls. The performance of the model not only surpasses that of other classical networks but also more focuses on characteristic regions relevant to Parkinson's pathology, thereby providing accurate and reliable diagnostic outcomes. |
