Automated Midline Shift Detection in Head CT Using Localization and Symmetry Techniques Based on User-Selected Slice.
Authors
Affiliations (3)
Affiliations (3)
- From the Department of Radiology (N.E.B., V.H., N.S., A.H., Z.Z., J.S.), Memorial Sloan Kettering Cancer Center, New York, New York [email protected].
- Department of Aerospace Engineering (H.S.), Indian Institute of Technology Madras, Chennai, India.
- From the Department of Radiology (N.E.B., V.H., N.S., A.H., Z.Z., J.S.), Memorial Sloan Kettering Cancer Center, New York, New York.
Abstract
Midline shift (MLS) is an intracranial pathology characterized by the displacement of brain parenchyma across the skull's midsagittal axis, typically caused by mass effect from space-occupying lesions or traumatic brain injuries. Prompt detection of MLS is crucial, because delays in identification and intervention can negatively impact patient outcomes. The gap we have addressed in this work is the development of a deep learning algorithm that encompasses the full severity range from mild to severe cases of MLS. Notably, in more severe cases, the mass effect often effaces the septum pellucidum, rendering it unusable as a fiducial point of reference. We sought to enable rapid and accurate detection of MLS by leveraging advances in artificial intelligence (AI). Using a cohort of 981 patient CT scans with a breadth of cerebral pathologies from our institution, we manually chose an individual slice from each CT scan primarily based on the presence of the lateral ventricles and annotated 400 of these scans for the lateral ventricles and skull-axis midline by using Roboflow. Finally, we trained an AI model based on the You Only Look Once object detection system to identify MLS in the individual slices of the remaining 581 CT scans. When comparing normal and mild cases to moderate and severe cases of MLS detection, our model yielded an area under the curve of 0.79 with a sensitivity of 0.73 and specificity of 0.72 indicating our model is sensitive enough to capture moderate and severe MLS and specific enough to differentiate them from mild and normal cases. We developed an AI model that reliably identifies the lateral ventricles and the cerebral midline across various pathologies in patient CT scans. Most importantly, our model accurately identifies and stratifies clinically significant and emergent MLS from nonemergent cases. This could serve as a foundational element for a future clinically integrated approach that flags urgent studies for expedited review, potentially facilitating more timely treatment when necessary.