TY - JOUR
T1 - Automatic quantification of ischemic injury on diffusion-weighted MRI of neonatal hypoxic ischemic encephalopathy
AU - Murphy, Keelin
AU - van der Aa, Niek E.
AU - Negro, Simona
AU - Groenendaal, Floris
AU - de Vries, Linda S.
AU - Viergever, Max A.
AU - Boylan, Geraldine B.
AU - Benders, Manon J. N. L.
AU - Išgum, Ivana
PY - 2017
Y1 - 2017
N2 - A fully automatic method for detection and quantification of ischemic lesions in diffusion-weighted MR images of neonatal hypoxic ischemic encephalopathy (HIE) is presented. Ischemic lesions are manually segmented by two independent observers in 1.5 T data from 20 subjects and an automatic algorithm using a random forest classifier is developed and trained on the annotations of observer 1. The algorithm obtains a median sensitivity and specificity of 0.72 and 0.99 respectively. F1-scores are calculated per subject for algorithm performance (median = 0.52) and observer 2 performance (median = 0.56). A paired t-test on the F1-scores shows no statistical difference between the algorithm and observer 2 performances. The method is applied to a larger dataset including 54 additional subjects scanned at both 1.5 T and 3.0 T. The algorithm findings are shown to correspond well with the injury pattern noted by clinicians in both 1.5 T and 3.0 T data and to have a strong relationship with outcome. The results of the automatic method are condensed to a single score for each subject which has significant correlation with an MR score assigned by experienced clinicians (p < 0.0001). This work represents a quantitative method of evaluating diffusion-weighted MR images in neonatal HIE and a first step in the development of an automatic system for more in-depth analysis and prognostication.
AB - A fully automatic method for detection and quantification of ischemic lesions in diffusion-weighted MR images of neonatal hypoxic ischemic encephalopathy (HIE) is presented. Ischemic lesions are manually segmented by two independent observers in 1.5 T data from 20 subjects and an automatic algorithm using a random forest classifier is developed and trained on the annotations of observer 1. The algorithm obtains a median sensitivity and specificity of 0.72 and 0.99 respectively. F1-scores are calculated per subject for algorithm performance (median = 0.52) and observer 2 performance (median = 0.56). A paired t-test on the F1-scores shows no statistical difference between the algorithm and observer 2 performances. The method is applied to a larger dataset including 54 additional subjects scanned at both 1.5 T and 3.0 T. The algorithm findings are shown to correspond well with the injury pattern noted by clinicians in both 1.5 T and 3.0 T data and to have a strong relationship with outcome. The results of the automatic method are condensed to a single score for each subject which has significant correlation with an MR score assigned by experienced clinicians (p < 0.0001). This work represents a quantitative method of evaluating diffusion-weighted MR images in neonatal HIE and a first step in the development of an automatic system for more in-depth analysis and prognostication.
UR - https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=85010931982&origin=inward
UR - https://www.ncbi.nlm.nih.gov/pubmed/28180081
U2 - https://doi.org/10.1016/j.nicl.2017.01.005
DO - https://doi.org/10.1016/j.nicl.2017.01.005
M3 - Article
C2 - 28180081
SN - 2213-1582
VL - 14
SP - 222
EP - 232
JO - NeuroImage. Clinical
JF - NeuroImage. Clinical
ER -