**3. Results**

The time interval between CEM and DCE-MRI was 2.5 days as a median value (range 1–16 days).

Table 2 reports the diagnostic performance of significant textural parameters for DCE-MRI and for dual-energy CEM in all views (i.e., CC, early and late MLO view), expressed in terms of AUC and *p*-value. The best result, considering the single feature in a univariate approach, was reached by the energy, range and GLN\_GLRLM extracted on DCE-MRI volume with an AUC of 0.72.

**Table 2.** Accuracy of significant textural parameters for DCE-MRI and for dual-energy CEM CC, early and late MLO view.


Figure 1 shows ROC curve trends of significant textural features: variance, correlation and IQR for mammography CC projection, kurtosis and skewness for mammography early-MLO projection and range, energy, entropy, GLN\_GLRLM and GLN\_GLSZM for DCE-MRI images.

Figure 2 shows the boxplots related to the above-mentioned parameters, to separate benign from malignant lesions.

Table 3 reports the performance achieved by the best classifiers designed to discriminate between benign and malignant lesions using CEM and DCE-MRI images.

The best performance, considering the CC view (ACC = 0.71; SENS = 0.71; SPEC = 0.71; PPV = 0.71; NPV = 0.71; AUC = 0.77), was reached with an SVM trained with 10-fold CV and balanced data (with SASYNO function) and a subset of four features (by LASSO and λminMSE). The subset of four robust textural features includes IQR, VARIANCE, CORRELATION and RLV.

**Figure 1.** ROC curve trends of significant textural features for DCE-MRI and for dual-energy CEM CC, early and late MLO view.

**Figure 2.** Boxplots of significant textural features for DCE-MRI and for dual-energy CEM CC, early and late MLO view.


**Table 3.** Performance of the best classifiers designed to discriminate between benign and malignant lesions.


**Table 3.** *Cont.*

The best performance, considering the early-MLO view (ACC = 0.76; SENS = 0.65; SPEC = 0.87; PPV = 0.82; NPV = 0.74; AUC = 0.73), was reached with an LDA trained with 10-fold CV and trained with balanced data (with ADASYN function) and all 48 textural features.

The best performance, considering the late-MLO view (ACC = 0.75; SENS = 0.71; SPEC = 0.77; PPV = 0.72; NPV = 0.75; AUC = 0.80), was reached with an LDA trained with 10-fold CV and balanced data (with ADASYN function) and a subset of 17 features (by LASSO). The subset of 17 robust textural features includes MEAN, MODE, MAD, RANGE,

#### VARIANCE, CONTRAST, CORRELATION, SRLGE, LRLGE, RLV, SZE, SZLGE, SZHGE, GLV\_GLSZM, BUSYNESS, COMPLEXITY and STRENGTH.

The best performance, considering all three mammographic projections (ACC = 0.79; SENS = 0.75; SPEC = 0.81; PPV = 0.78; NPV = 0.79; AUC = 0.81), was reached with an NNET trained with LOOCV and balanced data (with ADASYN function) and a subset of 15 features (by LASSO). The subset of 15 robust textural features includes IQR, VARIANCE, CORRELATION, LRHGE, GLV\_GLRLM and SZLGE among textural features extracted from CC view; MODE, CONTRAST and GLV\_GLRLM among textural features extracted from early-MLO view; MODE, STD, RANGE, IQR, CORRELATIOND and COMPLEXITY among textural features extracted from late-MLO view.

With regard to DCE-MRI images, the best performance (ACC = 0.74; SENS = 0.73; SPEC = 0.75; PPV = 0.74; NPV = 0.73; AUC = 0.72) was reached with an SVM trained with 10-fold CV and balanced data (with SASYNO function) and a subset of 15 features (by LASSO and λminMSE). The subset of 15 robust textural features includes MODE, MEDIAN, STD, MAD, ENTROPY, SUM AVERAGE, SRE, GLN\_GLRLM, SRHGE, LZE, ZSN, ZP, LZHGE, GLV\_GLSZM and ZSV.

Table 4 reports the performance achieved by the best classifiers to discriminate benign from malignant lesions when features extracted from CEM and DCE-MRI were combined. The best results overall (ACC = 0.84; SENS = 0.73; SPEC = 0.92; PPV = 0.90; NPV = 0.79; AUC = 0.88) were obtained considering a subset of 18 textural features extracted from all three mammographic views (CC, early MLO and late MLO) and DCE-MRI with an LDA trained with 10-fold CV and with balanced data (with ADASYN function). The subset of 18 robust textural features (by LASSO and λminMSE) includes IQR, VARIANCE, CORRELATION, LRHGE, GLV\_GLRLM and RLV among textural features extracted from CC mammographic view; MODE and CONTRAST among textural features extracted from early-MLO mammographic view; STD, RANGE, CORRELATION and COMPLEXITY among textural features extracted from late-MLO mammographic view; RANGE, KUR-TOSIS, AUTOCORRELATION, LRHGE, LZE and GLV\_GLSZM among textural features extracted from DCE-MRI images.


**Table 4.** Performance achieved by the best classifiers to discriminate between benign and malignant lesions for combined CEM and DCE-MRI.

Figure 3 shows the ROC curves of the best classifier obtained combining features from CEM and DCE-MRI.

**Figure 3.** LDA classifier ROC curve trained with 18 robust radiomic features from CEM and DCE-MRI.
