Combining different views of mammographic texture resemblance (MTR) marker of breast cancer risk
Research output: Contribution to conference › Conference abstract for conference › Research › peer-review
Standard
Combining different views of mammographic texture resemblance (MTR) marker of breast cancer risk. / Sun, S.; Karemore, Gopal; Chernoff, Konstantin; Karssemeijer, N.; Nielsen, Mads.
2011. Abstract from 5th international workshop on densiometry and breast cancer risk assessment, San Francisco, United States.Research output: Contribution to conference › Conference abstract for conference › Research › peer-review
Harvard
APA
Vancouver
Author
Bibtex
}
RIS
TY - ABST
T1 - Combining different views of mammographic texture resemblance (MTR) marker of breast cancer risk
AU - Sun, S.
AU - Karemore, Gopal
AU - Chernoff, Konstantin
AU - Karssemeijer, N.
AU - Nielsen, Mads
PY - 2011
Y1 - 2011
N2 - PURPOSEMammographic density is a well established breast cancer risk factor. Texture analysis in terms of the Mammographoc Texture Resemblance (MTR) marker has recently shown to add to risk segregation. Hitherto only single view MTR analysis has been performed. Standard mammography examinations include RMLO, RCC, LMLO, LCC views. Thus here we investigated the interrelation and combination of MTR scoring from several views.METHOD AND MATERIALSThe study included mammograms of 495 women (aged 58.0±5.7 years) from the Dutch screening program of which 250 controls were without diagnosis the subsequent 4 years whereas 245 cases had a diagnosis 2-4 years post mammography.We employed the MTR supervised texture learning framework to perform risk evaluation from a single mammography view. In the framework 20,000 pixels were sampled and classified by a kNN pixel classifier. A feature selection step is included to reduce input space dimensionality. Weak local decision scores for pixels were fused into an overall risk score.The dataset was randomly separated into a training data set (60%) and a test data set (40%).Risk scores for combinations of views were obtained by linear and quadratic discriminant analysis (LDA, QDA) where respectively Fisher criterion and Likelihood ratio were used as combination scores. LDA and QDA parameters were obtained from the training set.Performance was evaluated by AUC statistics. Correlations were analyses as Pearson’s linear correlation coefficient.RESULTSNo significant difference in age was found between cases and controls.The AUC values for RMLO, LMLO, RCC and LCC views are respectively 0.604, 0.579, 0.602 and 0.605. Combination of views yielded RMLO & LMLO: 0.600; RCC & LCC: 0.612; RMLO & RCC: 0.632; LMLO & LCC: 0.623.The correlation of scores from contralateral views was 0.72-0.75. Scatter plots are shown below.CONCLUSIONThe MTR AUCs are a little lower than earlier reported probably due to the smaller training set.MTR scores obtained from two contralateral views correlated well, but not as highly as previously reported on density (>0.85). We conclude that view combination may reduce some of the risk
AB - PURPOSEMammographic density is a well established breast cancer risk factor. Texture analysis in terms of the Mammographoc Texture Resemblance (MTR) marker has recently shown to add to risk segregation. Hitherto only single view MTR analysis has been performed. Standard mammography examinations include RMLO, RCC, LMLO, LCC views. Thus here we investigated the interrelation and combination of MTR scoring from several views.METHOD AND MATERIALSThe study included mammograms of 495 women (aged 58.0±5.7 years) from the Dutch screening program of which 250 controls were without diagnosis the subsequent 4 years whereas 245 cases had a diagnosis 2-4 years post mammography.We employed the MTR supervised texture learning framework to perform risk evaluation from a single mammography view. In the framework 20,000 pixels were sampled and classified by a kNN pixel classifier. A feature selection step is included to reduce input space dimensionality. Weak local decision scores for pixels were fused into an overall risk score.The dataset was randomly separated into a training data set (60%) and a test data set (40%).Risk scores for combinations of views were obtained by linear and quadratic discriminant analysis (LDA, QDA) where respectively Fisher criterion and Likelihood ratio were used as combination scores. LDA and QDA parameters were obtained from the training set.Performance was evaluated by AUC statistics. Correlations were analyses as Pearson’s linear correlation coefficient.RESULTSNo significant difference in age was found between cases and controls.The AUC values for RMLO, LMLO, RCC and LCC views are respectively 0.604, 0.579, 0.602 and 0.605. Combination of views yielded RMLO & LMLO: 0.600; RCC & LCC: 0.612; RMLO & RCC: 0.632; LMLO & LCC: 0.623.The correlation of scores from contralateral views was 0.72-0.75. Scatter plots are shown below.CONCLUSIONThe MTR AUCs are a little lower than earlier reported probably due to the smaller training set.MTR scores obtained from two contralateral views correlated well, but not as highly as previously reported on density (>0.85). We conclude that view combination may reduce some of the risk
M3 - Conference abstract for conference
Y2 - 8 June 2011 through 9 June 2011
ER -
ID: 33552145