Skip to main content
. 2013 Jun 20;20(6):1067–1075. doi: 10.1136/amiajnl-2012-001503

Table 4.

Predictive accuracy of WEKA-generated classifiers for predicting readers’ decisions to report a mammographic location as suspicious based on image and gaze features

Model Selected features Classifier ROC area
Group Model_LoCo Gaze+Contrast+Correlation+Energy+μ22+μ31 Bayesian Network 0.900±0.022
Group Model_LoRo_R1 Gaze+Contrast+Entropy+Correlation+μ12+μ22 RBF 0.929±0.050
Group Model_LoRo_R2 Gaze+Contrast+Entropy+Energy+μ22+μ31 Bayesian Network 0.888±0.056
Group Model_LoRo_R3 Gaze+Correlation+Contrast+Energy+μ12+μ22+σ44 RBF 0.872±0.056
Group Model_LoRo_R4 Gaze+Correlation+Energy+ μ22+μ31 Bayesian Network 0.919±0.051
Group Model_LoRo_R5 Gaze+Contrast+Correlation+Energy+μ12+μ22 Adaboost w/ MLP 0.907±0.051
Group Model_LoRo_R6 Gaze+Contrast+Correlation+Energy+μ12+μ22+σ32 MLP 0.808±0.081
Individual_Model_R1 Gaze Adaboost w/ MLP 0.927±0.050
Individual_Model_R2 All features MLP 0.864±0.061
Individual_Model_R3 Dwell+Returns NaiveBayes 0.766±0.082
Individual_Model_R4 All features MLP 0.799±0.073
Individual_Model_R5 All features MLP 0.919±0.046
Individual_Model_R6 All features MLP 0.891±0.067

Gaze, dwell+initial+returns; LoCo, leave-one-case-out; LoRo, leave-one-reader-out; RBF, radial basis function; ROC, receiver operating characteristics.