IDENTIFICATION OF PARKINSON'S DISEASE BIOMARKERS IN A MULTI-MODAL AI FRAMEWORK USING RAW MRI AND CLINICAL DATA

The primary objective of this study was to construct a 3D Convolutional Neural Network (CNN) based on multimodal MRI, clinical and demographic data. The purpose was to effectively discriminate healthy controls and individuals diagnosed with Parkinson's Disease (PD) in advanced stages. Subsequently, the designed model was used to develop another CNN, mirroring the architecture of the previous one. This secondary CNN was developed to differentiate between controls and PD patients in earlier disease stages. The ultimate goal of this project was to design a CNN capable of predicting the progression of PD, classifying whether an individual's prognosis would remain stable or worsening. This predictive capacity was achieved by merging MRI data with patients’ clinical and demographic information. Three different cohorts of PD patients and controls have been selected from different scanners. Firstly, 148 PD patients (86 early, 62 severe) and 60 controls were recruited. The second cohort included 56 early PD patients and 20 controls from PPMI database. Lastly, the third cohort included 91 early PD patients (de-novo) and 38 controls. All participants underwent an MRI scan at baseline and a clinical evaluation at baseline and after 2 years of follow-up. For the third aim of the study, PD patients stable from those who experiencing worsening were identified using a k-means clustering based on baseline and follow-up UDPRS-III value. CNN, which are mathematical representations of the human neural architecture with multiple hidden layers of artificial neurons, were applied. No pre-processing on MRI images were integrated into the pipeline. To effectively reduce the input dimensionality, a volumetric patch was extracted from the total MRI image including brainstem and basal ganglia areas and used as input in the CNN. In each experiment, the available dataset is partitioned into training, validation, and test set. With the transfer learning technique, knowledge extracted in easier tasks was used as initialization in more specific tasks with the aim of facilitating features extraction. Data augmentation (including elastic deformation, cropping, flipping, and scaling of the images) was applied to augment dataset's dimensionality by synthesizing new volumes. In our comprehensive analysis, CNN model demonstrated the capability to effectively differentiate PD subjects from controls, achieving a good level of performance. Notably, we observed different accuracy into distinguishing PD patients based on the disease stage relative to controls. Considering moderate-severe PD patients relative to controls, the accuracy rate on the test dataset reached nearly 75%, only relying on MRI data for classification. However, considering early-stage PD versus controls, the accuracy rate was around 50% on the test set, pointing out certain challenges in the extraction of discriminative features during the initial stages of the disease. Leveraging on transfer learning, we noted an improvement in test accuracy, which increased to approximately 65%. In this study, we successfully developed a 3D CNN using MRI data that exhibited encouraging results in distinguishing between controls and PD. By integrating MRI data with clinical and demographic information, our CNN demonstrated promising results and offers a valuable tool for early diagnosis and personalized treatment planning for PD patients.