Automated Segmentation of the Substantia Nigra in 3D MRI Images

”Parkinson’s Disease (PD) is the second most common neurodegenerative disorder” (Rizek et al., 2016), affecting over 6 million people worldwide (Armstrong & Okun, 2020). It primarily impacts motor functions and can also cause speech disorders, memory problems and depression (Sveinbjornsdottir, 2016). The Substantia Nigra (SN), located in the midbrain, is critically affected by PD due to the degeneration of dopamine-producing cells (Latif et al., 2021). Monitoring SN degeneration is crucial for understanding disease progression. Traditionally, magnetic resonance imaging (MRI) has been used to detect PD, with manual segmentation of the SN to assess cell loss (Heim et al., 2017). However, manual segmentation is time-consuming, prone to operator bias and subject to inter- and intra-observer variability. Automated segmentation using convolutional neural networks (CNNs) offers significant advantages by reducing manual workload and providing more consistent results (Heye et al., 2013). 3D imaging, in particular, enhances the understanding of anatomical structures by capturing inter-slice relationships, which 2D CNNs cannot do. 3D CNNs can learn these inter-slice relationships, leading to more accurate segmentation of volumetric medical data like MRI (Shamsudeen et al., 2022). In this study, I successfully utilized a large dataset by implementing semi-supervised learning techniques, self-training and consistency regularization. The research focused on performing automated 3D segmentation of the substantia nigra (SN) using three-dimensional MRI volumes. This approach aimed to enhance both the accuracy and efficiency of detecting and monitoring the progression of PD. The results demonstrated that consistency regularization with minimal augmentation yielded the best performance in SN segmentation. In contrast, self-training did not produce good results, as it relied solely on the labeled data. Pseudo-labels were not incorporated due to their poor quality, which stemmed from the model’s initial inability to accurately segment the SN. Future work could focus on further refining the consistency regularization approach, particularly by experimenting with different augmentation strategies and hyper-parameter settings. By optimizing these factors, it is possible to enhance segmentation accuracy, which would ultimately aid in the early detection and monitoring of PD´s progression.