Tumor Masking and Radiomics Feature Extraction

Two experienced breast radiologists blinded to pathological outcomes manually delineated the tumor region of interest (ROI) by outlining the tumor margin on the low-energy and recombined images with standard CC projection before NAC via the ITK-SNAP software, as shown in Figure 1 . If contradictory, other senior radiologists will evaluate the tumor mask again to reach agreement. The recombined images were used as reference to determine the tumor boundary on the low-energy images. Radiomics features per patient were then extracted from pretreatment CESM images with manually segmented ROIs. The task of radiomics feature extraction was conducted in the AK software (Artificial Intelligence Kit; GE Healthcare, China, Shanghai).

Example of delineating region of interest (ROI) in a 35 year-old woman with a 4.5-cm mass in the left breast. (Left) The low-energy and (Right) recombined images with cranial caudal (CC) projection.

To ensure reproducibility of radiomics feature extraction, we employed intra-class correlation coefficients (ICCs) for assessing the intra- and inter-observer agreement of ROI delineation. First, two radiologists with 8 years (Reader 1) and 9 years (Reader 2) of experience in diagnosis of breast cancer delineated the ROI of 30 randomly chosen CESM images. One week later, Reader 1 repeated the same procedure. An ICC > 0.75 was considered as substantial agreement.