Joeky Senders
117 Automating clinical chart review agreement increased, model performance according to the AUC increased as well. All six variables that were labelled with a near perfect interrater agreement ( κ between 0.8 and 1) were classified with an AUC above 0.95, whereas this performance was achieved for only two out of nine variables with a κ lower than 0.8. 0.0 0.2 0.4 0.6 0.8 1.0 False Positive Rate 0.0 0.2 0.4 0.6 0.8 1.0 True Positive Rate A ROC curves for laterality characteristics Left (AUC = 0.984 ± 0.017) Right (AUC = 0.973 ± 0.024) Multifocality (AUC = 0.816 ± 0.048) 0.0 0.2 0.4 0.6 0.8 1.0 False Positive Rate 0.0 0.2 0.4 0.6 0.8 1.0 True Positive Rate B ROC curves for locational characteristics Frontal lobe (AUC = 0.96 ± 0.015) Temporal lobe (AUC = 0.965 ± 0.022) Parietal lobe (AUC = 0.966 ± 0.022) Occipital lobe (AUC = 0.982 ± 0.019) Corpus callosum (AUC = 0.945 ± 0.031) 0.0 0.2 0.4 0.6 0.8 1.0 False Positive Rate 0.0 0.2 0.4 0.6 0.8 1.0 True Positive Rate C ROC curves for tumor aspect Necrosis (AUC = 0.962 ± 0.023) Cystic (AUC = 0.957 ± 0.056) Ring enhancement (AUC = 0.926 ± 0.029) Heterogenous enhancement (AUC = 0.853 ± 0.049) 0.0 0.2 0.4 0.6 0.8 1.0 False Positive Rate 0.0 0.2 0.4 0.6 0.8 1.0 True Positive Rate D ROC curves for other characteristics Hemorrhage (AUC = 0.901 ± 0.052) Edema (AUC = 0.949 ± 0.026) Mass effect (AUC = 0.899 ± 0.037) FIGURE 1. Receiver operating characteristic curves for all extracted radiological characteristics, grouped by domain including A) tumor laterality B) tumor location C) tumor aspect and D) other characteristics.
Made with FlippingBook
RkJQdWJsaXNoZXIy ODAyMDc0