Premium
CT‐based brain segmentation and volumetry using deep learning methods
Author(s) -
Srikrishna Meera,
Pereira Joana B,
Heckemann Rolf A,
Volpe Giovanni,
Zettergren Anna,
Kern Silke,
Westman Eric,
Skoog Ingmar,
Schöll Michael
Publication year - 2020
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.045824
Subject(s) - medicine , white matter , magnetic resonance imaging , radiology , deep learning , nuclear medicine , grey matter , segmentation , artificial intelligence , neuroimaging , computer science , psychiatry
Background Computed tomography (CT) is the most commonly used brain examination tool for the initial assessment of neurodegenerative diseases such as dementia disorders. It is widely available, affordable and provides short scan times, however, it is mainly used for visual assessment of brain integrity and exclusion of co‐pathologies while magnetic resonance imaging (MRI) is the preferred modality for the extraction of regional brain volumetric measures, obtained by brain segmentation into tissue classes and anatomical regions. We developed an automated, deep learning‐based approach to segment CT scans into intracranial volume (ICV), grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF). Method 139 CT and T1‐weighted MRI scans acquired on the same day in participants of the Gothenburg H70 Birth Cohort Study (70% female) were included in this study. SPM12 was used to extract GM, WM and CSF tissue class labels and Pincram for creating ICV masks from the MRIs. These labels were used as training input and as the reference criteria for subsequent volumetric and similarity comparisons. In a training subcohort (n=120), CT scans were co‐registered to their corresponding MRI. A U‐Net deep learning model was developed using the Keras software package which underwent two training sessions: 1. Predict ICVs in CT and extract brain; 2. Predict GM, WM and CSF volumes with extracted CT images as input, preserving the predictions in individual space. MRI labels were incorporated during the training sessions including 120 datasets. Result Nineteen validation datasets which were not included in the training were used for the prediction process without MRI intervention. Continuous Dice scores between ICV, GM, WM and CSF predictions from the model in the validation datasets and independently MRI‐derived labels were 0.97, 0.79, 0.81 and 0.72 with a volumetric Pearson correlation coefficient, r of 0.93, 0.84, 0.73 and 0.71, respectively. Conclusion Our pilot study developed a deep learning model that enables the automated segmentation and volumetric measure extraction from cranial CT. With further optimization, training on larger and more diverse datasets, we aim at improving the model to execute predictions comparable to MRI‐based approaches, thus, developing a diagnostic support tool based on routinely acquired CT scans.