TY - CHAP
T1 - Evaluating classifiers for atherosclerotic plaque component segmentation in MRI
AU - van Engelen, Arna
AU - Bruijne, Marleen De
AU - Schneider, Torben
AU - van Dijk, Anouk C.
AU - Kooi, M. Eline
AU - Hendrikse, Jeroen
AU - Nederveen, Aart
AU - Niessen, Wiro J.
AU - Botnar, Rene
PY - 2017
Y1 - 2017
N2 - Segmentation of tissue components of atherosclerotic plaques in MRI is promising for improving future treatment strategies of cardiovascular diseases. Several methods have been proposed before with varying results. This study aimed to perform a structured comparison of various classifiers, training set sizes, and MR image sequences to determine the most promising strategy for methodology development. Five different classifiers (linear discriminant classifier (LDC), quadratic discriminant classifier (QDC), random forest (RF), and support vector classifiers with both a linear (SVMlin) and radial basis function kernel (SVMrbf)) were evaluated. We used carotid MRI data from 124 symptomatic patients, scanned in 4 centres with 2 different MRI protocols (45 and 79 patients). Firstly, learning curves of accuracy as a function of increasing training data size showed stabilisation of performance after using ∼10–15 patients for training. Best results were found for LDC, QDC and RF. Intraplaque haemorrhage was most accurately classified in both protocols, and lowest accuracy was found for the lipid-rich necrotic core. Secondly, for LDC and RF it was shown that leaving out different MRI sequences usually negatively affects results for one or more classes. However, leaving out T2-weighted scans did not have a big impact. In conclusion, several classifiers obtain generally good results for classification of plaque components in MRI. Identification of intraplaque haemorrhage is the most promising, and lipid-rich necrotic core remains the most difficult.
AB - Segmentation of tissue components of atherosclerotic plaques in MRI is promising for improving future treatment strategies of cardiovascular diseases. Several methods have been proposed before with varying results. This study aimed to perform a structured comparison of various classifiers, training set sizes, and MR image sequences to determine the most promising strategy for methodology development. Five different classifiers (linear discriminant classifier (LDC), quadratic discriminant classifier (QDC), random forest (RF), and support vector classifiers with both a linear (SVMlin) and radial basis function kernel (SVMrbf)) were evaluated. We used carotid MRI data from 124 symptomatic patients, scanned in 4 centres with 2 different MRI protocols (45 and 79 patients). Firstly, learning curves of accuracy as a function of increasing training data size showed stabilisation of performance after using ∼10–15 patients for training. Best results were found for LDC, QDC and RF. Intraplaque haemorrhage was most accurately classified in both protocols, and lowest accuracy was found for the lipid-rich necrotic core. Secondly, for LDC and RF it was shown that leaving out different MRI sequences usually negatively affects results for one or more classes. However, leaving out T2-weighted scans did not have a big impact. In conclusion, several classifiers obtain generally good results for classification of plaque components in MRI. Identification of intraplaque haemorrhage is the most promising, and lipid-rich necrotic core remains the most difficult.
UR - http://www.scopus.com/inward/record.url?scp=85022181570&partnerID=8YFLogxK
U2 - 10.1007/978-3-319-60964-5_14
DO - 10.1007/978-3-319-60964-5_14
M3 - Other chapter contribution
AN - SCOPUS:85022181570
SN - 9783319609638
VL - 723
T3 - Communications in Computer and Information Science
SP - 156
EP - 168
BT - Medical Image Understanding and Analysis - 21st Annual Conference, MIUA 2017, Proceedings
PB - Springer Verlag
T2 - 21st Annual Conference on Medical Image Understanding and Analysis, MIUA 2017
Y2 - 11 July 2017 through 13 July 2017
ER -
