C3DKeras for Parkinson-related Dementia Identification

The diagnosis of Parkinson's disease dementia (PDD) still follows the one-year rule, which may be too late for the optimal treatment. Since up to 83% of Parkinson’s disease (PD) patients eventually develop dementia later in the disease progression, early identification would allow for timely administration of the appropriate treatment, potentially leading to an increased life expectancy. Additionally, as there is an exponential growth of the aging population in the world the number of people suffering from PD and consequently of PDD is increasing. However, until now no therapeutic method has been discovered for completely treat the subjects affecting by this type of neurodegenerative disease. Therefore, the early detection could be the only way to increase the life expectancy. This thesis provided a systematic literature review on PDD identification through Machine Learning (ML) algorithms and a novel methodology for the identification of PDD subjects from MRI scans. A particular 3D CNN, called C3DKeras, pre-trained on large video-clips on sports, was used as core part of the proposed pipeline, whereas the entire methodology was conducted under the structure of ablation studies. Specifically, four ablation experiments were performed, in which a new part was added to the previous one in the pipeline as follows: Experiment 0, no modification from the original C3DKeras; Experiment 1, addition of weighting class; Experiment 2, addition of unfrozen layers; and Experiment 3, addition of rigorous hyperparameter tuning. All the experiments were conducted by taking into account two datasets downloaded from Parkinson’s Progression Markers Initiative (PPMI), namely PDD and prodromal PD. Experiment 3, with an AUC of 86%, SE of 100%, ACC of 73%, and F1-score of 73% for the PDD class; achieved the highest performance among the four experiments and among the literature research results. Therefore, as the entire methodology was conducted under the structure of ablation studies, it was evident how each fine-tuning step significantly improve the overall performance. Hence, this study provides an innovative approach for the identification and classification of PDD subjects as well as those with PD using the particular CNN called C3DKeras