Patient dataset:

In this study, 74 patients aged between 36 and 69 years were undergone to a CESM exam at I.R.C.C.S. Istituto Tumori "Giovanni Paolo II" of Bari (Italy) between February 2016 and July 2017. For these patients, the inclusion criterion was to have a positive result confirmed by histology. Then, each case was classified according to BPE classification (ACR BIRADS) [8]. Primary and secondary (when present) lesions were detected and their diameters were measured: a total of 86 lesions of 0.5-10 cm in size, of which 13 benign and 73 malignant after the histological test, were evaluated. ROIs with same size were extracted from lesions, BPE and no-contrasted background, both on LE and RC images (Figs. 1 and 2, respectively).

Data acquisition:

For this purpose, a modified DM device derived from a Senographe Essential (GE Healthcare) was used. The image acquisition started two minutes after the iodinated CM injection (1.5 ml/kg of body mass of Visipaque 320 mg I/ml), acquiring low- (at 26-30 kVp) and high-energy (at 45-49 kVp) images of both breasts. In this phase, craniocaudal (CC) and mediolateral oblique (MLO) views were obtained for each type of acquisition. The acquisition process was completed within 5 minutes recombined images were obtained from LE and HE images by dual-subtraction technique. All of images were in DICOM format.

Data analysis:

First, 4 statistical features of gray-levels, such as mean, standard deviation, maximum and minimum values, were extracted. Then, two additional parameters (coefficient and range of variance) were computed, obtaining a total of 12 features for each ROI (both LE and RC types). Finally, a state-of-the-art classifier, such as Random Forest (RF) classifier, was implemented in order to discriminate BPE regions from ROIs including lesions.

Each value of the 12 features extracted from the BPE regions and ROIs with lesions were normalized with respect to the corresponding value extracted from the non-contrasted background region of the same patient.

The performance of the proposed model was validated on 100 rounds of 10-fold cross-validation in terms of Area Under the Curve (AUC)  of the Receiver Operating Characteristic (ROC) curve, accuracy, sensitivity and specificity, giving the median value and the interquantile range (IQR) for each performance descriptor. A further evaluation of the performance was also given for each BPE class, considering moderate and marked classes as a single category, labelled high BPE class, because of the reduced number of patients belonging to these categories.