3D MRI Volume Segmentation Using 2D U-Net Models: A Focus on Deep Brain Structures of the Striatum for Parkinson's Disease early diagnosis

Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms, affecting millions of people worldwide. It results from the loss of dopamine in critical brain structures such as the substantia nigra and the striatum. Nowadays, several research studies highlight the importance of medical imaging techniques for early PD diagnosis. Among these techniques, Magnetic Resonance Imaging (MRI) holds a central position in neuroimaging, offering unparalleled insights into deep brain structures. Nevertheless, accurate segmentation of complex brain regions, such as the striatum, remain essential for precise diagnoses and decision-making. In this study, we present a novel method that utilizes 2D U-Net models to partition 3D MRI volumes and automatically segment the striatum, avoiding multi-class segmentation complexities. Our dataset comprised 97 MRI volumes with their respective segmentation. Each 3D MRI volume was converted into 3112D images slices of128x128 dimensions. In our methodology, we applied two distinct data preprocessing techniques to generate optimal inputs for our U-N et model. We proposed both manual and automatic selection method to reduce the number of slices by filtering out homogeneous and redundant ones. Comprehensive evaluation metrics, including Dice Loss and Binary Cross Entropy, were employed to ensure rigorous assessment. Results demonstrated promising potential with filtered slices yielding better striatum segmentation and a training accuracy of 90.59%.