Skip to main content

Table 4 Prediction of masking

From: Using automated texture features to determine the probability for masking of a tumor on mammography, but not ultrasound

Method MSE AUC NRI Reclassification
Correctly upwards Correctly downwards
Nulla 0.0682 (0.0095) 0.500 (0.000)    
Clinical findingsb 0.0657 (0.0085) 0.734 (0.037)    
Univariate selectionc 0.0656 (0.0085) 0.743 (0.036) 27.9 (16.2) 57.9 (9.1) 56.1 (3.0)
Lassoc 0.0655 (0.0084) 0.747 (0.036) 33.1 (15.6) 60.0 (8.7) 56.6 (3.0)
Boostingc 0.0654 (0.0084) 0.747 (0.036) 32.5 (15.5) 59.8 (8.6) 56.5 (3.0)
Random forestc 0.0656 (0.0087) 0.739 (0.035) 4.4 (16.3) 45.1 (9.0) 57.1 (3.4)
Observed PMDd 0.0645 (0.0082) 0.753 (0.036) 35.7 (14.4) 58.5 (8.2) 59.4 (2.9)
  1. Summary statistics (mean and standard deviation) of MSE, AUC, and the net reclassification improvement (NRI) in percentages obtained from logistic regression models with clinical predictors and the observed or predicted PMD using various regression methods. All measurements were obtained by 3-fold cross-validation with 100 repetitions
  2. AUC area under the curve, BMI body mass index, HRT hormone replacement therapy, MSE mean squared error, NRI net reclassification improvement, PMD percentage mammographic density
  3. aLogistic regression model without any predictors
  4. bLogistic regression model with clinical predictors (age, BMI, prior breast surgery, menopausal and HRT status, imaging technique) but without PMD
  5. cLogistic regression model with clinical predictors and PMD predicted from texture features using univariate selection, lasso, boosting, or random forest
  6. dLogistic regression model with clinical predictors and the original PMD values (“observed PMD”)