Automated Semisupervised Measurement of Optic Nerve Sheath Diameter From CT Following Traumatic Brain Injury.
Authors
Affiliations (7)
Affiliations (7)
- Gilbert S. Omenn Department of Computational Medicine and Bioinformatics, University of Michigan, Ann Arbor, Michigan, USA, umich.edu.
- Department of Radiology, Michigan Medicine, Ann Arbor, Michigan, USA, umich.edu.
- Department of Emergency Medicine, Michigan Medicine, Ann Arbor, Michigan, USA, umich.edu.
- Max Harry Weil Institute for Critical Care Research and Innovation, Michigan Medicine, Ann Arbor, Michigan, USA, umich.edu.
- Division of Neurocritical Care, Department of Neurosurgery, University of Michigan, Ann Arbor, Michigan, USA, umich.edu.
- Michigan Institute for Data and AI in Society (MIDAS), University of Michigan, Ann Arbor, Michigan, USA, umich.edu.
- Center for Data-Driven Drug Development and Treatment Assessment (DATA), University of Michigan, Ann Arbor, Michigan, USA, umich.edu.
Abstract
Automated measurement of optic nerve sheath diameter (ONSD) from computed tomography (CT) scans is clinically important, but deep learning methods are often limited by the scarcity of high-quality, expert-labeled data given the labor-intensive nature of manual segmentations. To address this, we present a fully automated modular pipeline combining deep learning and semisupervised learning techniques for ONSD estimation from axial head CT scans. The pipeline features three stages: (1) deep learning-based slice selection, (2) semisupervised machine learning for optic nerve segmentation utilizing both labeled and unlabeled publicly available data, and (3) geometric and morphological ONSD measurement. When validated on public and internal datasets, our approach generalized better than traditional supervised segmentation methods, achieving an intersection over union (IoU) of 0.561 ± 0.141 and Dice coefficient of 0.707 ± 0.131 on previously unseen data. Slice-wise measurement accuracy varied based on measurement distance from the ocular globe, yielding mean absolute errors as low as 1.779 and 1.899 mm for the right and left ONSD, respectively. Our findings highlight the potential for semisupervised deep learning to deliver fully automated ONSD measurements and the framework's adaptability to difficult medical imaging tasks even with limited, low-quality ground truth for training.