The atrophy and iron load of the subthalamic nucleus (STN), substantia nigra (SN), and red nucleus (RN) characterize neurological disorders such as Parkinson's disease. Despite their key role in motor control and cognition, they have been overlooked partly due to challenges in imaging these small and deep-seated midbrain structures and lack of accurate and efficient segmentation methods. Quantitative susceptibility mapping (QSM), an MRI technique to estimate in-vivo iron contents, provides a distinguishable contrast of these iron-rich structures. We propose an automated segmentation approach for demarcating STN, SN and RN, based on a deep convolutional neural networks model with U-Net architecture and MRI data. MRI data including T1-weighted, FLAIR, QSM and relaxometry (R2star maps) in addition to manually delineated nuclei masks on QSM images of 40 individuals were used to train a deep learning model in single-modal (using each MRI modality individually) and multi-modal (combinations of MRI data) setups. Five-fold cross validation results revealed that the multi-modal model using QSM and FLAIR performed best (average Dice scores of 0.84, 0.91 and 0.94 for STN, SN and RN, respectively). Subsequently, the best model was applied on a dataset including 208 adults (age range 20-80). Cross-sectional analyses showed significant associations between iron content and age in the STN (r=0.47, p<0.001), SN (r=0.28, p<0.001) and RN (r=0.34, p<0.001). Our automatic segmentation approach using deep neural networks offers a novel tool to access and accurately evaluate volume and iron load of these small midbrain nuclei which can lead to a deeper understanding of their function. Authors:
Farshad Falahati, Karolinska Institutet, Sweden
Jonatan Gustavsson, Karolinska Institutet, Sweden
Alireza Salami, Karolinska Institutet, Sweden
Grégoria Kalpouzos, Karolinska Institutet, Sweden

About the Presenter(s)
Dr. Falahati is a research engineer at Karolinska Institutet, Sweden. He has a PhD in Neuroscience and an MSc in Medical Engineering. His research interests include the use of neuroimaging and machine learning techniques to study the human brain.

Connect on Linkedin
https://www.linkedin.com/in/farshad-falahati/

See this presentation on the full schedule – Friday Schedule

Bookmark