*3.1. Experimental Setting*

#### 3.1.1. Dataset: INbreast

The INbreast dataset has been widely used in previous studies [18,28,50,51] and was one of the first established datasets of full-field digital mammograms (FFDM) acquired in 2011 at Centro Hospitalar de S. Joo, Breast Centre, Porto [52]. A total of 410 images were extracted with 115 abnormal lesion cases ranging from mass, calcification, and architectural distortions, with both craniocaudal (CC) and mediolateral oblique (MLO) views. Subsequently, the extracted images were exclusively updated by the authors with permission, along with the annotated ground truth range of interest (ROI) of the segmented mass region. Note that 112 mass images were included for this study that ranges across four breast density classifications, further classified based on their mass types: benign and malignant. To avoid sampling bias, 80% of the images were randomly selected for training, with the remaining 20% used for testing and validation for all stages, and were independent of the breast mass types and density level. Finally, augmentation settings were set into degrees of rotation of 30◦ to 300◦, horizontally flipped, and scaled to randomized 1.0 to 1.3 scale factor. Augmentation settings that alter the hue, contrast, brightness, and saturation were excluded to avoid unintentional intensity changes affecting the breast density.
