| Popis: |
Parkinson's disease (PD) is the second most prevalent neurological disorder, predominantly affecting older people. With no existing cure, the early detection of PD, where symptoms are not entirely evident but indicative of the disease's onset, is critical. This study aims to design and develop AI-based diagnostic methods that can detect these early signs of PD with high accuracy, thereby facilitating more effective disease management. This study focuses on developing a method that not only identifies PD at an early stage but also provides clinicians with a tool to interpret the decisions taken by the AI models to avoid misdiagnosis. In this study, a T2-weighted 3D Magnetic resonance imaging (MRI) dataset is used to analyze detailed morphological, textural, and structural changes. The MRI scans are pre-processed using brain extraction, image registration, bias correction, normalization, and segmentation processes. Upon segmentation, feature extraction was applied to the segmented subcortical regions using radiomics tools, resulting in the extraction of 107 features. The top 20 features were selected through Pearson’s correlation, recursive feature elimination, and a ranking model, which are responsible for the ML model’s class prediction. Statistical validation of these features was also performed using Analysis of Variance (ANOVA), pairwise t-tests, and Kruskal-Wallis H-tests to ensure that the identified 20 features were dominant for the prediction. Based on the identified features, several Machine Learning (ML) models were used to identify the best classifier for the provided real-world MRI scans. The Gradient Boosting (GB) algorithm achieved better prediction accuracy among the compared models. Incorporating the Synthetic Minority Over-sampling Technique (SMOTE) to address data imbalances significantly improved the model's performance, boosting accuracy to 96.8 % from 87 %. Further, multiple Explainable Artificial Intelligence (XAI) techniques were deployed to enhance the transparency and interpretability of the models. These techniques provide insights into how each identified feature influences predictions by the classifier, assisting clinicians in making trustworthy decisions when planning diagnosis and treatment measures. |
