TY - GEN
T1 - Contusion segmentation from subjects with traumatic brain injury
T2 - 2014 IEEE 11th International Symposium on Biomedical Imaging, ISBI 2014
AU - Rao, A.
AU - Ledig, C.
AU - Newcombe, V.
AU - Menon, D.
AU - Rueckert, D.
N1 - Publisher Copyright:
© 2014 IEEE.
PY - 2014/7/29
Y1 - 2014/7/29
N2 - Traumatic Brain Injury (TBI) occurs when a sudden injury causes trauma to the brain. Contusions are one of the most common types of lesion that arise after TBI, and they can be observed on a subject's MRI or CT. Since it is hypothesised that indices such as contusion load may be potential biomarkers for TBI, the ability to segment contusions is highly desirable. Currently, we are not aware of any fully automated methods that address this segmentation task. In this paper we present a completely automated random-forest based approach to contusion segmentation that uses multi-modality MRI. Given a training set of MR images and ground-truth segmentations, a set of features is derived for each voxel that describe both the local neighbourhood and longer-range contextual information in the images. A random forest is trained using these features and the ground-truth voxel labels, and used to produce an automatic contusion segmentation of an unseen test subject. We evaluate the method using 6-fold cross-validation on a dataset consisting of 23 subjects, obtaining a mean DICE overlap of 0.60.
AB - Traumatic Brain Injury (TBI) occurs when a sudden injury causes trauma to the brain. Contusions are one of the most common types of lesion that arise after TBI, and they can be observed on a subject's MRI or CT. Since it is hypothesised that indices such as contusion load may be potential biomarkers for TBI, the ability to segment contusions is highly desirable. Currently, we are not aware of any fully automated methods that address this segmentation task. In this paper we present a completely automated random-forest based approach to contusion segmentation that uses multi-modality MRI. Given a training set of MR images and ground-truth segmentations, a set of features is derived for each voxel that describe both the local neighbourhood and longer-range contextual information in the images. A random forest is trained using these features and the ground-truth voxel labels, and used to produce an automatic contusion segmentation of an unseen test subject. We evaluate the method using 6-fold cross-validation on a dataset consisting of 23 subjects, obtaining a mean DICE overlap of 0.60.
UR - http://www.scopus.com/inward/record.url?scp=84922452524&partnerID=8YFLogxK
U2 - 10.1109/isbi.2014.6867876
DO - 10.1109/isbi.2014.6867876
M3 - Conference contribution
AN - SCOPUS:84922452524
T3 - 2014 IEEE 11th International Symposium on Biomedical Imaging, ISBI 2014
SP - 333
EP - 336
BT - 2014 IEEE 11th International Symposium on Biomedical Imaging, ISBI 2014
PB - Institute of Electrical and Electronics Engineers Inc.
Y2 - 29 April 2014 through 2 May 2014
ER -